An AI-based multi-factory server intelligent production scheduling method and system

Abstract

The application discloses an AI-based multi-factory server intelligent production scheduling method and system, comprising the following steps: each factory edge end compresses production states into low-dimensional representation vectors and extracts local rules through a preference perception information bottleneck encoder; a central server calculates implicit preference vectors and implicit constraint conditions of each factory from the low-dimensional representation vectors and historical trajectories by using maximum entropy inversion reinforcement learning; global rule bases are generated by generalizing and conflict resolving the local rules based on the implicit preference vectors by using inductive logic programming; global rule base decisions are preferentially matched during online scheduling, and a neural network is called when there is no match or low confidence, and successful decisions are converted into new rules which are added to the rule base after being confirmed; meanwhile, the compression coefficient is dynamically adjusted according to preference reasoning uncertainty to control the communication code rate. The application reduces communication load, realizes interpretable scheduling decisions, aligns implicit preferences and safely migrates cross-factory rules.

Description

Technical Field

[0001] This invention belongs to the field of intelligent manufacturing and artificial intelligence technology, and in particular relates to an AI-based multi-factory server intelligent production scheduling method and system. Background Technology

[0002] Multi-factory collaborative production scheduling is one of the core issues in the field of intelligent manufacturing. In existing technologies, centralized scheduling methods upload real-time status data from each factory (such as equipment operating status, buffer backlog, order progress, etc.) to a central server. The server then runs optimization algorithms or scheduling models to make unified decisions and distributes instructions to each factory for execution. Research in this area includes traditional operations research methods based on mixed-integer programming, as well as data-driven methods based on deep reinforcement learning (such as multi-agent deep deterministic policy gradients, multi-agent proximal policy optimization, etc.). To protect the data privacy of each factory, federated learning frameworks have been introduced into multi-factory scheduling scenarios. Each factory trains its model locally and only uploads model parameters or gradients; the central server updates the global model through aggregation. Furthermore, rule-based scheduling systems are widely used in industrial settings. Managers summarize their experience into heuristic rules (such as "shortest processing time priority" and "earliest delivery date priority"), and the system quickly generates scheduling instructions through rule matching. In recent years, graph neural networks have also been used to model process dependencies within factories and material flow relationships between factories to capture high-order correlation features of scheduling states.

[0003] However, existing technologies still have the following shortcomings. First, in terms of communication efficiency, centralized scheduling that uploads all raw data results in high bandwidth consumption, making it difficult to adapt to network-constrained environments; while simple feature filtering or downsampling lacks theoretical optimality and may discard crucial scheduling information. Second, in terms of interpretability, deep reinforcement learning models output black-box decisions, which are difficult for on-site managers to understand and trust; static rule systems, although interpretable, are difficult to adapt to dynamically changing environments and cannot be migrated across factories. Third, in terms of preference alignment, the actual optimization goals of each factory (such as potentially prioritizing orders from a strategic customer) and implicit constraints (such as equipment maintenance periods not recorded in the system) are often not explicitly declared in the system, causing a misalignment between the central server's scheduling based on a unified reward function and the actual intentions of the factories, leading to execution deviations or even constraint violations. Fourth, in terms of rule reuse, excellent scheduling experience between factories cannot be safely transferred in symbolic form; new factories need to train from scratch or manually write rules. To address the above problems, this invention proposes an AI-based multi-factory server intelligent production scheduling method and system. Summary of the Invention

[0004] To overcome the aforementioned shortcomings of existing technologies, this invention provides an AI-based intelligent production scheduling method and system for multiple factory servers, which solves the problems of high communication load, black box decision-making, misaligned preference constraints, and inability to migrate rules across factories in existing multi-factory scheduling technologies.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A multi-factory server-based intelligent production scheduling method and system based on AI includes the following steps:

[0007] S1: At the edge of each factory, the original data of the current production status is compressed into a low-dimensional representation vector through the preference-aware information bottleneck encoder; and symbolic scheduling rules are extracted from the local policy network to form a local rule set;

[0008] S2: After receiving the low-dimensional representation vector and the local rule set, the central server uses maximum entropy inversion reinforcement learning to calculate the latent preference vector and latent constraints of each factory from the low-dimensional representation vector and historical scheduling trajectory.

[0009] S3: The central server, based on the implicit preference vector of each factory, uses inductive logic programming to generalize the local rule set and resolve conflicts in the generalized rules to generate a global rule base;

[0010] S4: During online scheduling, the central server prioritizes matching the rules in the global rule base to generate scheduling decisions based on the low-dimensional representation vectors of each factory. If no matching rule exists or the confidence of the matching rule is lower than a preset threshold, the main scheduling neural network is invoked to generate scheduling decisions. In response to the invocation of the main scheduling neural network and the execution feedback data meeting the preset good conditions, a new rule is extracted from the corresponding state-action pair and added to the global rule base after confirmation.

[0011] Preferred options also include:

[0012] The central server calculates the compression coefficient of the preference-aware information bottleneck encoder at the edge of each factory based on the uncertainty metric value of preference inference for each factory, and sends the compression coefficient to the corresponding factory edge to control the uplink communication code rate; the compression coefficient is negatively correlated with the uncertainty metric value; wherein, the uncertainty metric value is a normalized value with a value range of 0-1; the compression coefficient has a value range of 0.01-1.00.

[0013] Preferably, the preset threshold is a confidence threshold with a value range of 0.5-0.7; the preset favorable condition is that the sum of the immediate reward values ​​of each factory after the scheduling decision is executed is higher than 20%-50% of the average reward value of the alternative decisions in the scheduling cycle, and no implicit constraint condition is triggered.

[0014] Preferably, the implicit constraints are obtained through constraint inverse reinforcement learning and are represented as a set of constraints extracted from the region boundary of state-action pairs that have not appeared in the historical trajectory; the implicit preference vector is a linear weight of the reward function, used to characterize the degree of importance the factory attaches to each scheduling objective.

[0015] Preferably, the generalization process includes:

[0016] Calculate the correlation coefficient between the predicate of each rule in the local rule set and each feature component in the latent preference vector. Replace the constants in the predicates with correlation coefficients greater than a preset generalization threshold with variables and add type constraints to generate general rules. The preset generalization threshold ranges from 0.3 to 0.6.

[0017] The conflict resolution includes: when two rules output different actions under the same conditions, a decision is made based on the similarity between the implicit preference vectors of the source factories of the two rules and the confidence of the two rules; wherein, the similarity is cosine similarity, with a value range of 0-1; when the similarity is greater than a first similarity threshold, a conditional predicate is added for differentiation, with the first similarity threshold ranging from 0.6 to 0.8; when the similarity is less than a second similarity threshold, the rule with higher confidence is adopted, with the second similarity threshold ranging from 0.2 to 0.4.

[0018] Preferably, an AI-based multi-factory server intelligent production scheduling system includes:

[0019] Multiple factory edge modules and one central server module;

[0020] The factory edge module includes:

[0021] An encoding unit is used to compress the original production state into a low-dimensional representation vector through a preference-aware information bottleneck encoder. The encoding unit includes an adjustable compression coefficient and a preference alignment coefficient.

[0022] Policy units are used to generate actions via a local policy network;

[0023] A rule extraction unit is used to extract symbolic rules from the strategy unit to form a local rule set;

[0024] The first communication unit is used to send the low-dimensional representation vector and the local rule set to the central server module, and to receive the scheduling decision and compression coefficient issued by the central server module.

[0025] The central server module includes:

[0026] The preference reasoning unit is used to receive the low-dimensional representation vectors and historical scheduling trajectories of all factories, and output the implicit preference vector, implicit constraints and uncertainty metric of preference reasoning for each factory. The uncertainty metric is a normalized value with a range of 0-1.

[0027] A bitrate control unit, connected to the preference inference unit, is used to calculate the compression coefficient of each factory based on the uncertainty metric value, wherein the compression coefficient ranges from 0.01 to 1.00.

[0028] The rule generalization and resolution unit, connected to the preference reasoning unit, is used to receive the implicit preference vector and the local rule set of all factories, perform generalization processing and conflict resolution on the local rule set, and output the global rule base.

[0029] The scheduling decision unit includes a rule matching subunit and a neural network subunit. The rule matching subunit receives the global rule base and the low-dimensional representation vectors of each factory, and outputs the matching rule and confidence level. The neural network subunit is activated when there is no matching rule or the confidence level is lower than a preset threshold, and outputs a scheduling decision. The preset threshold ranges from 0.5 to 0.7.

[0030] The rule evolution unit, connected to the scheduling decision unit, is used to receive the scheduling decision and execution feedback data output by the neural network subunit and output new rules;

[0031] The second communication unit is paired with the first communication unit to exchange data. It is used to send scheduling decisions and compression coefficients to the edge modules of each factory, and to receive the low-dimensional representation vector and the local rule set.

[0032] Preferably, the central server module further includes a constraint verification unit, which is connected to the scheduling decision unit and the preference reasoning unit respectively. After the scheduling decision unit outputs a scheduling decision, it verifies whether the scheduling decision violates the implicit constraint of any factory. If it does, it triggers the scheduling decision unit to regenerate the scheduling decision.

[0033] Preferably, the rule evolution unit further includes a confirmation interface for visually displaying the logical structure of the new rule and receiving confirmation, modification, or rejection instructions.

[0034] Preferably, the preference inference unit also periodically sends the implicit preference vector and uncertainty metric value obtained by the current inference to each factory edge module through the second communication unit; the encoding unit receives the implicit preference vector and uncertainty metric value and adjusts its preference alignment coefficient accordingly; the preference alignment coefficient ranges from 0.1 to 1.0.

[0035] Preferably, the central server module further includes a data storage unit for persistently storing the historical scheduling trajectories, low-dimensional representation vectors, global rule base, and rule evolution logs of all factories; the storage retention period set for the data storage unit is 30-365 days.

[0036] The technical effects and advantages of the AI-based multi-factory server intelligent production scheduling method and system of this invention are as follows:

[0037] 1. This invention deploys a preference-aware bottleneck encoder at the edge of each factory, compressing the high-dimensional raw production state into a low-dimensional representation vector before uploading it to a central server. The encoder's training objective includes mutual information maximization, compression regularization, and preference interpretability terms, ensuring that it prioritizes retaining features related to scheduling decisions and factory preferences during compression, while discarding redundant information. Compared to the traditional centralized scheduling method of directly uploading all raw data, this invention significantly reduces the amount of uplink communication data, lowers network bandwidth consumption, and enables the system to operate stably in network-constrained industrial environments. Simultaneously, the adaptive bitrate control module dynamically adjusts the compression coefficient based on the uncertainty of preference inference in each factory, allocating higher bitrates to factories with drastic state changes and lower bitrates to stable factories, thus optimizing the allocation of communication resources while ensuring scheduling quality.

[0038] 2. This invention extracts symbolic scheduling rules from a local policy network through a rule extraction unit, and generalizes and resolves these rules using inductive logic programming to form a global rule base. During online scheduling, rule matching is prioritized for decision-making; only when there is no matching rule or the rule confidence falls below a preset threshold does it revert to the neural network. This hybrid decision-making mechanism allows most scheduling decisions to be presented as "if-then" symbolic rules, enabling factory managers to understand the generation logic of each scheduling instruction. Compared to pure black-box neural network scheduling methods, this invention significantly improves the system's transparency and auditability, reducing the deployment resistance of artificial intelligence systems in real-world industrial scenarios.

[0039] 3. This invention deploys an inverted preference inference module on a central server, employing maximum entropy inverted reinforcement learning to deduce latent preference vectors from the historical scheduling trajectories of each factory, while simultaneously using constraint inverse reinforcement learning to extract latent constraints. This method does not rely on the factory's pre-declared objective function, but rather learns the true optimization preferences from actual behavior (e.g., a factory claims to minimize delays but actually prioritizes orders from a key customer) and unrecorded hard constraints (e.g., a certain piece of equipment is prohibited from scheduling during specific periods). Global scheduling decisions are optimized based on these mined preferences and constraints, avoiding scheduling mismatches and constraint violations caused by misaligned objectives. Compared to traditional methods using a uniformly preset reward function, this invention can adaptively align with the true intentions of each factory, improving the actual executability of the scheduling scheme and factory satisfaction.

[0040] 4. This invention employs inductive logic programming to generalize the local rule set. The generalization direction is guided by the implicit preference vectors of each factory: predicates highly correlated with preferences are replaced with variables and type constraints are added to generate general rules; constants with low correlation to preferences are retained as factory-specific. This preference-guided generalization mechanism allows effective rules from a factory, proven in practice, to be safely transferred symbolically to other factories with similar preferences, while avoiding the erroneous generalization of irrelevant factory details. The rule generalization and resolution unit also resolves cross-factory rule conflicts by comparing the preference similarity and rule confidence of the rule source factories, ensuring the consistency and reliability of the global rule base. Compared to traditional methods where each factory independently maintains its scheduling rules and cannot share them, this invention enables the cross-factory replication of excellent management experience, reducing the cold start cost of scheduling models in new factories.

[0041] 5. This invention deeply integrates information bottleneck compression, inverted preference reasoning, and neural symbolic rule learning to form a closed loop: the encoder's compression coefficient is dynamically adjusted based on the uncertainty of preference reasoning; the rule generalization direction is guided by the latent preference vector; successful decisions of the neural network can be automatically converted into symbolic rules and added to the rule base; the evolution results of the rule base in turn feed back to influence subsequent preference re-reasoning. This closed loop enables the system's scheduling performance, communication efficiency, and interpretability to be continuously optimized synchronously, with each module's learning mutually reinforcing rather than operating independently. Compared with the existing design where compression, reasoning, and scheduling are separated, this invention has stronger adaptive capabilities and long-term evolutionary capabilities.

[0042] 6. The constraint verification unit of this invention verifies whether any implicit constraints of any factory are violated before the scheduling decision is issued. If a violation is found, a re-decision is triggered to avoid executing the violation. The rule evolution unit has a confirmation interface, allowing engineers to visually review, modify, and confirm newly extracted rules. This integrates automated learning with human experience verification to ensure the quality and security of the rule base. The data storage unit persistently manages historical trajectories, preference results, and rule logs, supporting periodic model retraining and incremental updates to the rule base, enabling the system to adapt to concept drift in the production environment. Attached Figure Description

[0043] Figure 1 This is a flowchart of an AI-based multi-factory server intelligent production scheduling method and system proposed in this invention. Detailed Implementation

[0044] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0045] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include," "contain," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "includes..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes the element.

[0046] refer to Figure 1This invention provides an AI-based intelligent production scheduling method and system for multiple factories, aiming to solve the technical problems of high communication load, uninterpretable decision-making, difficulty in aligning implicit preferences and constraints, and inability to migrate rules across factories in multi-factory collaborative scheduling. The method of this invention includes: each factory edge end compresses the raw data of the current production state into a low-dimensional representation vector through a preference-aware information bottleneck encoder, and extracts symbolic scheduling rules from the local policy network to form a local rule set; after receiving the low-dimensional representation vector and the local rule set, the central server uses maximum entropy inversion reinforcement learning to calculate the latent preference vector and latent constraint conditions of each factory from the low-dimensional representation vector and historical scheduling trajectories; then, based on the latent preference vector of each factory, the local rule set is generalized using inductive logic programming, and conflict resolution is performed on the generalized rules to generate a global rule base; during online scheduling, the central server prioritizes matching the rules in the global rule base to generate scheduling decisions based on the current low-dimensional representation vector of each factory. If there is no matching rule or the confidence of the matching rule is lower than a preset threshold, the main scheduling neural network is invoked to generate a scheduling decision. In response to the invocation of the main scheduling neural network and the execution feedback data meeting the preset good conditions, a new rule is extracted from the corresponding state-action pair, confirmed, and added to the global rule base; at the same time, the central server dynamically adjusts the compression coefficient of each factory encoder according to the uncertainty metric of the preference inference of each factory to control the uplink communication code rate. This invention also provides a system for implementing the above method, comprising multiple factory edge modules and a central server module. These modules collaborate to complete state compression, preference reasoning, rule generalization, hybrid decision-making, and adaptive communication control. This invention effectively reduces the communication bandwidth consumption of multi-factory scheduling, achieves symbolic and interpretable output of scheduling decisions, automatically mines implicit preferences and constraints of each factory, supports secure generalization and reuse of cross-factory rules, and forms a closed-loop evolution mechanism, significantly improving the intelligence level and practical executability of multi-factory collaborative scheduling.

[0047] Example 1

[0048] Purpose of implementation:

[0049] This embodiment aims to demonstrate the complete scheduling process of the present invention in a multi-factory collaborative production environment, and to verify the effectiveness of state compression and rule extraction, preference reasoning, rule generalization, and hybrid scheduling decision-making.

[0050] Implementation System:

[0051] This embodiment uses the following system modules:

[0052] There are three factory edge modules (factory A, factory B, and factory C). Each factory edge module includes: an encoding unit, a strategy unit, a rule extraction unit, and a first communication unit.

[0053] A central server module includes: a preference inference unit, a bitrate control unit, a rule generalization and resolution unit, a scheduling decision unit (the scheduling decision unit includes a rule matching subunit and a neural network subunit), a rule evolution unit, a second communication unit, a constraint verification unit, and a data storage unit.

[0054] Implementation steps:

[0055] S1. State compression and rule extraction:

[0056] The encoding units at the edge of each factory compress the raw data of the current production state into a low-dimensional representation vector through a preference-aware information bottleneck encoder. In this embodiment, the raw state includes 128 features such as equipment operating status, number of buffers, and order urgency. The encoder is a three-layer fully connected neural network that outputs a 16-dimensional low-dimensional representation vector. The initial compression coefficient of the encoder is set to 0.5 (within the range of 0.01 to 1.00), and the initial preference alignment coefficient is set to 0.3 (within the range of 0.1 to 1.0). The training objective of the encoder consists of three terms: the first term maximizes the mutual information between the low-dimensional representation vector and the expected reward of the scheduling decision; the second term minimizes the mutual information between the low-dimensional representation vector and the raw state multiplied by the compression coefficient; and the third term maximizes the log-likelihood of the distribution predicted by the preference inference unit for the factory reward function multiplied by the preference alignment coefficient.

[0057] Each factory's strategy unit generates actions through a local policy network trained using the PPO algorithm. After every 1000 scheduling cycles, the rule extraction unit uses the C4.5 decision tree algorithm to extract symbolic scheduling rules from the policy network, with the decision tree depth limited to 4 levels, forming a local rule set. Each rule takes the form of:

[0058] Factory A Rule R_A1: If (Equipment A load is greater than 0.75) and (order slack time is less than 2 hours), then the action is an outsourcing request.

[0059] Factory B Rule R_B1: If (the temperature of equipment B is greater than 80 degrees Celsius), then the action is to reduce speed.

[0060] Each rule comes with a confidence score, which is equal to the proportion of samples covered by the corresponding leaf node of the decision tree.

[0061] S2. Preference Reasoning:

[0062] Each factory's first communication unit uploads its low-dimensional representation vector and local rule set to the central server. After receiving the data, the central server's second communication unit uses maximum entropy inversion reinforcement learning to calculate each factory's latent preference vector and latent constraints from the low-dimensional representation vector and historical scheduling trajectories (past 30 days).

[0063] The feature vector designed in this embodiment contains four components: order delay time, energy consumption, equipment changeover frequency, and inventory level. The core of maximum entropy inverted reinforcement learning is to find a reward function that maximizes the probability of observing historical scheduling trajectories under that reward function. The reward function is represented as a linear weighted sum of the feature vectors, and the weight vector is the latent preference vector. The solution process iteratively updates the weights through gradient ascent. In each iteration, the expected value of the features under the current policy needs to be calculated and compared with the expected value of the features in the actual trajectory. After 100 iterations, the latent preference vectors of each factory are obtained:

[0064] Factory A's latent preference vector is [0.65, 0.20, 0.10, 0.05], indicating that Factory A attaches great importance to order delays.

[0065] Factory B's latent preference vector is [0.10, 0.70, 0.15, 0.05], indicating that Factory B attaches great importance to energy consumption.

[0066] The latent preference vector of factory C is [0.30, 0.30, 0.30, 0.10], indicating that factory C is relatively balanced in its pursuit of various objectives.

[0067] Implicit constraints are obtained through constraint inverse reinforcement learning: a set of constraints is extracted from the boundaries of state-action pairs that have never appeared before using a single-class support vector machine (parameter ν is set to 0.05). Specifically, all state-action pairs in historical trajectories are treated as positive samples, and a hypersphere is trained to describe the boundaries of normal behavior. The region outside the hypersphere is the violation region, and its boundary is approximated by a system of linear inequalities. For example, the constraints for factory A include: equipment C is not allowed to perform scheduling actions during the working hours from 14:00 to 15:00.

[0068] S3. Rule generalization and conflict resolution:

[0069] The rule generalization and resolution unit generalizes the local rule set based on the latent preference vector of each factory using inductive logic programming. Inductive logic programming is a symbolic learning method that learns logical procedures from instances and background knowledge. Specifically, the correlation coefficient between each predicate in each rule and each feature component in the latent preference vector is calculated. Taking the rule "Equipment A load is greater than 0.75" of factory A as an example, the correlation coefficient between this predicate and the "order delay" feature in factory A's latent preference vector is 0.82. This value is greater than the preset generalization threshold of 0.5 (the preset generalization threshold ranges from 0.3 to 0.6), so the constant "Equipment A" in the predicate is replaced with the variable "a certain device", and the type constraint "this device is a temperature-sensitive device" is added to generate the general rule: "If the load of the temperature-sensitive device is greater than 0.75, then an outsourcing request is made."

[0070] The conflict resolution process is as follows: When Factory A's rule "Request outsourcing if equipment A load is greater than 0.75 and order slack time is less than 2 hours" and Factory B's rule "Suspend production if global load is greater than 85%" output "Request outsourcing" and "Suspend production" respectively under the same state (global load greater than 85%), the cosine similarity between the latent preference vectors of Factory A and Factory B is calculated to be 0.35. This similarity is less than the second similarity threshold of 0.4 (the second similarity threshold ranges from 0.2 to 0.4). Therefore, the confidence of the two rules is compared: the confidence of Factory A's rule is 0.92, and the confidence of Factory B's rule is 0.78. The rule with the higher confidence, namely Factory A's rule, is adopted. Finally, a global rule base is generated, containing 23 general rules.

[0071] S4. Hybrid Scheduling and Rule Evolution:

[0072] During online scheduling, the central server's scheduling decision unit prioritizes matching rules from the global rule base using the low-dimensional representation vectors of each factory. In the current state, rule number G7 is matched, with a confidence level of 0.88, which is greater than the preset threshold of 0.6 (the preset threshold ranges from 0.5 to 0.7). Therefore, the rule matching subunit directly outputs the scheduling decision: Factory A accelerates, Factory B outsources, and Factory C stands by.

[0073] If no matching rule exists or the confidence level of the matching rule is below a preset threshold, the neural network subunit is invoked to generate a scheduling decision. The neural network subunit adopts a Transformer structure and includes an 8-head attention mechanism. It takes the concatenation of the low-dimensional representation vectors of all factories as input and outputs the action probability distribution of each factory.

[0074] The scheduling decision is sent to the edge terminals of each factory via the second communication unit, and executed by the first communication unit. Feedback data is collected after execution: the sum of the immediate reward values ​​of each factory in this scheduling is higher than 28% of the average reward value of the alternative decisions during this scheduling cycle, and no implicit constraints are triggered. This 28% falls within the 20% to 50% range required by the preset excellent conditions, thus satisfying the preset excellent conditions. The rule evolution unit extracts a new rule from this state-action pair: "If the time difference between material arrivals across factories is less than 10 minutes and the load of equipment A in factory is greater than 0.7, then initiate joint assembly." After confirmation by engineers through the confirmation interface, this new rule is added to the global rule base.

[0075] Implementation results:

[0076] This embodiment successfully implemented multi-factory collaborative scheduling. The amount of communication data was compressed from the original 128 dimensions to 16 dimensions, a reduction to 12.5% ​​of the original data volume. Scheduling decisions were output in the form of symbolic rules, ensuring full interpretability. Automatic migration of rules across factories was successful. After 30 days of operation, the number of rules in the global rule base stabilized between 20 and 30, with no implicit constraint violations occurring.

[0077] Example 2

[0078] Purpose of implementation:

[0079] This embodiment aims to verify that when there are undeclared real targets in each factory, the inverted preference reasoning module can accurately mine implicit preference vectors and implicit constraints, thereby guiding subsequent scheduling decisions.

[0080] Implementation System:

[0081] This embodiment uses a factory edge module (factory D) and a central server module, and the module composition is the same as in embodiment 1.

[0082] Implementation steps:

[0083] S1. State compression and rule extraction:

[0084] The encoding unit at the edge of factory D compresses the original production state (including 128-dimensional features such as equipment load, order queue, and customer tags) into a 16-dimensional low-dimensional representation vector. The rule extraction unit extracts initial rules from the local policy network, such as "prioritize the order with the shortest relaxation time".

[0085] S2. Preference Reasoning:

[0086] The central server receives the low-dimensional representation vector and historical scheduling trajectory (5000 samples from the past 30 days) of factory D. The preference inference unit employs maximum entropy inversion reinforcement learning, and the feature vector includes: standard order processing time, order relaxation time, whether the order belongs to customer X (a binary variable), and equipment switchover time. The reward function weights are initialized to zero vectors and iterated 100 times. The output latent preference vector is as follows: the weight corresponding to standard processing time is 0.4, the weight corresponding to relaxation time is 0.3, the weight corresponding to customer X is 2.3, and the weight corresponding to switchover time is 0.5. The weight of 2.3 for customer X is significantly higher than that of other features, indicating that factory D's true preference is to prioritize processing orders from customer X, rather than nominally minimizing the total processing time.

[0087] Constraint-based inverse reinforcement learning was used to extract implicit constraints: analysis of historical trajectories revealed that factory D never scheduled equipment M for processing on Friday afternoons. Using a single-class support vector machine (parameter ν set to 0.05), the hypersphere boundary was obtained, and the constraint condition was derived: equipment M is not allowed to perform any processing tasks during the Friday afternoon period. It was verified that equipment M was used for special calibration during this period.

[0088] S3. Rule generalization and conflict resolution:

[0089] The rule generalization and resolution unit combines the rule "prioritize the order with the shortest relaxation time" with the high weight of customer X in the implicit preference vector, and generalizes it to "if the order customer is equal to X, then the priority is increased by 20%".

[0090] S4. Hybrid Scheduling and Rule Evolution:

[0091] During online scheduling, the central server automatically reserves resources for customer X's order in task allocation based on the mined implicit preference vector and implicit constraints, and avoids scheduling device M on Friday afternoons. After the execution feedback meets the preset favorable conditions, the rule evolution unit extracts a new rule: "If the waiting time for customer X's order is greater than 4 hours, then suspend other orders," which is then added to the global rule base after confirmation.

[0092] Implementation results:

[0093] This embodiment successfully uncovered the true preferences and implicit constraints of factory D. In subsequent scheduling, the overdue rate of customer X's orders decreased by 56%, and there were no further instances of irregular scheduling of equipment M. The uncertainty metric for preference inference decreased from an initial 0.8 to 0.2, indicating that the inference results are stable and reliable.

[0094] Example 3

[0095] Purpose of implementation:

[0096] This embodiment aims to verify that inductive logic programming can achieve safe generalization of cross-factory rules and resolve conflicts between multiple factory rules through a conflict resolution mechanism.

[0097] Implementation System:

[0098] This embodiment uses three factory edge modules (factory E, factory F, and factory G) and a central server module, with the same module composition as in embodiment 1.

[0099] Implementation steps:

[0100] S1. State compression and rule extraction:

[0101] The local extraction rules for each factory are as follows:

[0102] Factory E: If the temperature of the spraying equipment is greater than 40 degrees Celsius and the humidity is greater than 70%, then suspend production.

[0103] Factory F: If (spraying equipment temperature is greater than 42 degrees Celsius) and (humidity is greater than 65%), then switch to low speed mode.

[0104] Factory G: If (the temperature of the spraying equipment is greater than 38 degrees Celsius), then accelerate the fan.

[0105] S2. Preference Reasoning:

[0106] The preference inference unit outputs the implicit preference vectors for each factory:

[0107] The latent preference vector of factory E is [0.8 (quality), 0.1 (energy consumption), 0.1 (output)].

[0108] The latent preference vector of factory F is [0.2 (quality), 0.6 (energy consumption), 0.2 (output)].

[0109] The latent preference vector of factory G is [0.2 (quality), 0.2 (energy consumption), 0.6 (output)].

[0110] S3. Rule generalization and conflict resolution:

[0111] The rule generalization and resolution unit calculates the correlation coefficient between each predicate in each rule and each feature component in the latent preference vector. For the rules of factory E, the correlation coefficient between "temperature greater than 40 degrees Celsius" and the quality feature is 0.91, and the correlation coefficient between "humidity greater than 70%" and the quality feature is 0.85, both greater than the preset generalization threshold of 0.5 (the threshold ranges from 0.3 to 0.6). Therefore, the constant "40 degrees Celsius" is replaced with the variable "temperature threshold", and "70%" is replaced with the variable "humidity threshold". Type constraints "this equipment belongs to spraying equipment" and "this product belongs to products with high surface quality requirements" are added to generate a general rule: "If the temperature of the spraying equipment is greater than the temperature threshold and the humidity is greater than the humidity threshold, and the product quality requirement is high, then it is recommended to suspend production." The default values ​​for the temperature threshold and humidity threshold are 80% of the average values ​​of all factories.

[0112] The conflict resolution process is as follows: When the spraying equipment temperature is 41 degrees Celsius and the humidity is 72%, the aforementioned general rule (derived from factory E) and factory G's rule "accelerate the fan if the spraying equipment temperature is greater than 38 degrees Celsius" are triggered simultaneously. The cosine similarity between the latent preference vectors of factory E and factory G is calculated to be 0.32, which is less than the second similarity threshold of 0.4 (the second similarity threshold ranges from 0.2 to 0.4). Therefore, the confidence levels of the two rules are compared: the general rule has a confidence level of 0.96, and factory G's rule has a confidence level of 0.73. The rule with the higher confidence level (i.e., production is suspended) is adopted. At the same time, the system automatically adds the condition constraint "humidity less than 65%" to factory G's rule to avoid future conflicts.

[0113] After a week of operation, the global rule base has evolved from the initial 100 rules to 52 general rules, reducing the number of rules by 48%.

[0114] S4. Hybrid Scheduling and Rule Evolution:

[0115] Newly extracted rules are added to the global rule base after confirmation, and the rule base continues to evolve and be optimized.

[0116] Implementation results:

[0117] This embodiment enables rule migration across factories, allowing factory E's quality management experience to be successfully applied to other factories. The conflict resolution mechanism avoids decision-making confusion caused by contradictory rules, resulting in a streamlined rule base with enhanced versatility.

[0118] Example 4

[0119] Purpose of implementation:

[0120] This embodiment aims to verify that the adaptive bit rate control module can dynamically adjust the compression coefficient of each factory under limited bandwidth conditions, prioritize the communication quality of factories with high uncertainty, and keep the overall bandwidth usage controllable.

[0121] Implementation System:

[0122] This embodiment employs five factory edge modules (factories H, I, J, K, and L) and one central server module. Each factory's encoding unit contains an adjustable compression factor, and the central server contains a bitrate control unit. The maximum bandwidth is 100 Mbps.

[0123] Implementation steps:

[0124] S1. State compression and rule extraction:

[0125] Initially, the compression coefficient for each factory was set to 0.5, and the low-dimensional representation vector dimension was 16. Each factory normally performed state compression and rule extraction.

[0126] S2. Preference Reasoning:

[0127] The central server's preference inference unit calculates the normalized uncertainty metric for each factory. This metric is calculated by taking the trace of the covariance matrix of the latent preference vector estimates and dividing it by the dimension of the latent preference vector. Initially, the uncertainty metric for each factory is approximately 0.1.

[0128] The rate control unit calculates the compression factor for each factory based on the uncertainty metric. The formula is: compression factor equals 0.01 plus (1.00 minus 0.01) multiplied by (1 minus the uncertainty metric). When the uncertainty metric is 0.1, the compression factor is calculated to be 0.9; when the uncertainty metric is 0.9, the compression factor is calculated to be 0.1. The rate control unit then distributes the compression factor to the corresponding factory edge.

[0129] S3. Rule generalization and conflict resolution:

[0130] This step is the same as in Example 1 and is not affected by bitrate control.

[0131] S4. Hybrid Scheduling and Rule Evolution:

[0132] One day, Factory I experienced a drastic change in production status due to a production line change, causing its uncertainty metric to rapidly rise from 0.1 to 0.85. Upon detecting this change, the central server's rate control unit immediately reduced Factory I's compression factor from 0.9 to 0.12 and transmitted it via the second communication unit. Upon receiving the new compression factor, Factory I's encoding unit superimposed Gaussian noise with a mean of 0 and a variance equal to the compression factor multiplied by 0.1 when outputting the low-dimensional representation vector. Then, it performed uniform quantization with a step size equal to the compression factor multiplied by 0.5, and finally performed Huffman coding before uploading. The uploaded information increased from 45 bytes per cycle to 280 bytes. Five minutes later, the uncertainty metric decreased to 0.2, and the compression factor rebounded to 0.8.

[0133] The uncertainty metrics of other factories remained stable, maintaining a high compression state (compression coefficient close to 0.9), and the overall bandwidth usage was always kept below 80Mbps.

[0134] Implementation results:

[0135] In this embodiment, the high-uncertainty factory achieved a higher communication code rate and quickly converged to a stable state; the low-uncertainty factory maintained a low code rate, with an overall bandwidth utilization improvement of 40%, and no scheduling errors caused by communication delays occurred.

[0136] Example 5

[0137] Purpose of implementation:

[0138] This embodiment demonstrates the complete interaction process between all modules in the system in a chronological manner, verifying the system's collaborative working capability.

[0139] Implementation System:

[0140] This embodiment uses two factory edge modules (E1, E2) and one central server module, and the configuration of each module is the same as in embodiment 1.

[0141] Implementation steps:

[0142] Initialization phase (time T=0)

[0143] The encoding units of E1 and E2 load the pre-trained model with a compression factor of 0.5; the policy unit starts local training; the rule extraction unit runs once every 24 hours to extract the initial rule sets R1_local and R2_local.

[0144] The preference inference unit of the central server requests E1 and E2 to upload historical trajectories (past 7 days) through the second communication unit. After receiving the trajectories, the preference inference unit runs maximum entropy inversion reinforcement learning and outputs the latent preference vectors of factory E1 and factory E2, as well as the latent constraints.

[0145] The rule generalization and resolution unit receives the local rule sets of the two factories and the implicit preference vector, runs inductive logic programming, and generates a global rule base.

[0146] The central server sends the global rule base and the current compression coefficient back to the first communication units of E1 and E2 through the second communication unit.

[0147] Online scheduling cycle (time T = 1 second)

[0148] Step S1: The encoding unit of E1 senses the current state s1 and outputs a low-dimensional representation vector Z1; the encoding unit of E2 senses the current state s2 and outputs a low-dimensional representation vector Z2.

[0149] Step S1 (continued): The first communication units of E1 and E2 send Z1 and Z2 to the second communication unit of the central server.

[0150] Step S4: The rule matching subunit in the central server's scheduling decision unit matches (Z1, Z2) with the global rule base and finds rule number 12. The confidence level of this rule is 0.92, which is greater than the preset threshold of 0.6. Therefore, the rule matching subunit directly outputs the actions: E1 Process order X, E2 Prepare material Y.

[0151] The constraint verification unit checks whether the above actions violate the implicit constraints of factory E1 or factory E2, and the check result is passed.

[0152] The second communication unit of the central server sends the scheduling decision to the first communication units of E1 and E2.

[0153] E1 and E2 perform actions, collect execution feedback rewards r1 and r2, and store them locally.

[0154] The rule evolution unit is waiting for execution feedback. This scheduling did not call the neural network subunit, so no new rules are extracted.

[0155] Periodic evolution (time T = 1 hour)

[0156] Over the past week, the neural network sub-unit has been invoked a total of 50 times. The rule evolution unit extracted three new rule candidates from these calls and presented them to engineers in a visual format through a confirmation interface.

[0157] The engineer confirmed that two of the rules were valid, modified the third rule, and accepted it. The rule evolution unit then added the confirmed new rule to the global rule base.

[0158] The central server simultaneously broadcasts the updated global rule base to E1 and E2.

[0159] Model update (time T=7 days)

[0160] The preference reasoning unit recollects the scheduling trajectories of each factory over the past 7 days and updates the implicit preference vectors and implicit constraints.

[0161] The rule generalization and resolution unit regeneralizes all local rules based on the new implicit preference vector, eliminating rules with low contribution.

[0162] The data storage unit automatically transfers old tracks older than 90 days to cold storage media.

[0163] Implementation results:

[0164] This embodiment demonstrates the complete closed-loop operation of the system. All modules interact as expected, the average latency of scheduling decisions is less than 1 second, the rule base continuously evolves and optimizes, and data storage management is highly efficient. The confirmation interface provides a human-machine collaborative rule confirmation mechanism.

[0165] Comparative Example 1

[0166] Traditional centralized scheduling methods:

[0167] Purpose of implementation:

[0168] This comparative example is used to highlight the technical advantages of the present invention, and is compared with the conventional method under the same experimental environment as Example 1.

[0169] Implementation System:

[0170] The traditional centralized scheduling system is configured as follows: each factory uploads all its raw state data to the central server without any compression; the server runs a multi-agent deep deterministic policy gradient (MADDPG) model and directly outputs the actions of each factory; the system does not include preference inference unit, rule extraction unit, rule generalization and resolution unit, code rate control unit, or rule evolution unit.

[0171] Implementation steps:

[0172] This comparative example is run according to the following procedure:

[0173] Each factory uploads all raw state data (128 dimensions) to the central server every second.

[0174] The central server concatenates the states of the five factories and inputs them into the MADDPG model, which then outputs the actions of each factory.

[0175] The server distributes the actions to each factory for execution.

[0176] No rule extraction, preference mining, adaptive compression, or rule evolution are performed.

[0177] Implementation results:

[0178] The comparison results between Example 1 (the present invention) and the example under the same operating conditions in three factories for 30 days are as follows:

[0179] Regarding communication traffic: The average communication traffic of Embodiment 1 of this invention is 28 megabytes per day, while the average communication traffic of the comparative conventional method is 520 megabytes per day. The communication traffic of this invention is reduced by approximately 94.6%.

[0180] Regarding the interpretability of scheduling decisions: This invention outputs symbolic rules and preference vectors, making the scheduling decisions fully interpretable; the comparative invention outputs a black-box neural network, which cannot explain the basis for the decisions.

[0181] Regarding the matching degree of the actual target of the factory: the present invention automatically mines the actual target of each factory through preference reasoning, and the matching degree reaches 89%; the comparative proportion uses a uniform preset reward function, and the matching degree is only 61%.

[0182] Regarding the response time to abnormal events: the average response time of this invention is 1.2 seconds; the average response time of the comparative example is 4.5 seconds.

[0183] Regarding the ability to migrate rules across factories: This invention achieves a 78% success rate, meaning that local rules can be effectively applied to other factories after generalization; the comparison ratio does not have rule migration capabilities.

[0184] Regarding the number of implicit constraint violations: the present invention had 0 implicit constraint violations during 30 days of operation; the comparative example had 12 constraint violations due to the failure to discover implicit constraints, including a quality incident caused by the incorrect scheduling of equipment M in factory C on Friday afternoon.

[0185] Furthermore, in the comparative example, the implicit preference of factory B (emphasis on energy consumption) was not captured by the model, and the central server frequently ordered factory B to accelerate production, resulting in a 22% increase in actual energy costs. In contrast, the preference inference module in this invention accurately uncovered the true objectives of each factory after running for 3 days, and the constraint mining module identified the special time periods of equipment M, thus avoiding the above-mentioned problems.

[0186] Compared with Examples 1-5 and Comparative Example 1, the five embodiments of the present invention verify the effectiveness of the technical solution from different dimensions. Example 1 demonstrates a complete basic scheduling process. The original 128-dimensional state is compressed to 16 dimensions by a preference-aware information bottleneck encoder, reducing the amount of communication data to 12.5% ​​of the original. Maximum entropy inversion reinforcement learning successfully mines the latent preference vectors of the three factories, aligning the global scheduling with the actual goals of the factories. Inductive logic programming generalizes local rules into general rules, forming a global rule base after conflict resolution, reducing the number of rules from 100 to 52. In the hybrid scheduling mechanism, rule matching takes priority, and the neural network is only called when there are no rules or low confidence, making the scheduling decision fully interpretable. The rule evolution module automatically extracts new rules from successful decisions, which are added to the rule base after confirmation by engineers, forming a closed-loop optimization. Example 2 focuses on preference reasoning and implicit constraint mining. When there is an undeclared true objective of "customer X priority" in factory D, the weight of customer X in the preference vector output by inverted reinforcement learning reaches 2.3 (other features are less than 0.5), accurately identifying the true preference. Constraint inverse reinforcement learning extracts the implicit constraint "device M disabled on Friday afternoon" from the never-appeared state-action pairs. In subsequent scheduling, the order delay rate of customer X decreased by 56%, and the number of violation scheduling was reduced to zero. Example 3 details rule generalization and conflict resolution. The temperature and humidity rules of factory E are generalized to universal rules applicable to all spraying equipment, and multiple rule conflicts are resolved through similarity comparison and confidence adjudication, resulting in a 48% reduction in the rule base. Example 4 verifies the effectiveness of adaptive bitrate control. The compression coefficient of the high-uncertainty factory (U value increases from 0.1 to 0.85) decreases from 0.9 to 0.12, and the uploaded information volume increases from 45 bytes to 280 bytes, converging within 5 minutes. The low-uncertainty factory maintains high compression, with the overall bandwidth controlled within 80Mbps, improving utilization by 40%. Example 5 presents the complete interaction process of each module of the system in a time sequence manner. The average delay of scheduling decision is less than 1 second, and the mechanisms such as data storage, rule broadcasting, and regular model updates all operate as expected.

[0187] Compared to the comparative example (traditional centralized MADDPG scheduling), this invention has significant advantages in several key indicators. In terms of communication efficiency, this invention reduces the average daily communication traffic from 520MB to 28MB, a reduction of 94.6%, significantly lowering network bandwidth requirements, making it particularly suitable for industrial internet environments. Regarding interpretability, this invention outputs symbolic rules and implicit preference vectors, making the scheduling decision logic clear and auditable; the comparative example's black-box neural network cannot explain its decision-making basis at all, making it difficult for on-site managers to trust. In terms of preference alignment, this invention automatically mines the factory's real targets through inverted reinforcement learning, achieving a matching degree of 89%, while the comparative example uses a uniform preset reward function, achieving a matching degree of only 61%, resulting in factory B incurring an additional 22% energy consumption due to being required to accelerate production. Regarding anomaly response speed, this invention has an average response time of 1.2 seconds, superior to the comparative example's 4.5 seconds. In terms of rule reuse capability, this invention achieves a 78% effective migration rate of local rules across factories, allowing factory E's quality management experience to directly benefit other factories, while the comparative example lacks any rule migration capability. Regarding constraint satisfaction, the present invention exhibited zero implicit constraint violations during 30 days of operation, while the comparative example, due to the failure to uncover implicit constraints, experienced 12 violations, including a quality incident caused by an incorrect scheduling of equipment M on Friday afternoon. In summary, the present invention significantly outperforms traditional centralized scheduling methods in the field of multi-factory intelligent production scheduling in five aspects: communication efficiency, interpretability, preference alignment, rule reuse, and constraint satisfaction.

[0188] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of protection of the claims.

[0189] In conclusion, the above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A multi-factory server intelligent production scheduling method based on AI, characterized in that, Includes the following steps: S1: At the edge of each factory, the original data of the current production status is compressed into a low-dimensional representation vector through the preference-aware information bottleneck encoder; and symbolic scheduling rules are extracted from the local policy network to form a local rule set. S2: After receiving the low-dimensional representation vector and the local rule set, the central server uses maximum entropy inversion reinforcement learning to calculate the latent preference vector and latent constraints of each factory from the low-dimensional representation vector and historical scheduling trajectory. S3: The central server, based on the implicit preference vector of each factory, uses inductive logic programming to generalize the local rule set and resolve conflicts in the generalized rules to generate a global rule base; S4: During online scheduling, the central server prioritizes matching the rules in the global rule base to generate scheduling decisions based on the low-dimensional representation vectors of each factory; if there are no matching rules or the confidence of the matching rule is lower than a preset threshold, the main scheduling neural network is invoked to generate scheduling decisions. In response to the invocation of the main scheduling neural network and the execution feedback data meeting the preset favorable conditions, new rules are extracted from the corresponding state-action pairs and added to the global rule base after confirmation.

2. The AI-based multi-factory server intelligent production scheduling method as described in claim 1, characterized in that, Also includes: The central server calculates the compression coefficient of the preference-aware information bottleneck encoder at the edge of each factory based on the uncertainty metric of preference reasoning for each factory, and sends the compression coefficient to the corresponding factory edge to control the uplink communication code rate. The compression coefficient is negatively correlated with the uncertainty metric; wherein the uncertainty metric is a normalized value with a range of 0-1; and the compression coefficient has a range of 0.01-1.

00.

3. The AI-based multi-factory server intelligent production scheduling method as described in claim 1, characterized in that, The preset threshold is a confidence threshold, with a value range of 0.5-0.7; the preset favorable condition is that the sum of the immediate reward values ​​of each factory after the scheduling decision is executed is higher than 20%-50% of the average reward value of the alternative decisions in the scheduling cycle, and no implicit constraint condition is triggered.

4. The AI-based multi-factory server intelligent production scheduling method as described in claim 1, characterized in that, The implicit constraints are obtained through constraint inverse reinforcement learning and are represented as a set of constraints extracted from the region boundary of state-action pairs that have not appeared in the historical trajectory; the implicit preference vector is a linear weight of the reward function, used to characterize the factory's emphasis on each scheduling objective.

5. The AI-based multi-factory server intelligent production scheduling method as described in claim 1, characterized in that, The generalization process includes: Calculate the correlation coefficient between the predicate of each rule in the local rule set and each feature component in the latent preference vector. Replace the constants in the predicates with correlation coefficients greater than a preset generalization threshold with variables and add type constraints to generate general rules. The preset generalization threshold ranges from 0.3 to 0.

6. The conflict resolution includes: when two rules output different actions under the same conditions, a decision is made based on the similarity between the implicit preference vectors of the source factories of the two rules and the confidence of the two rules; wherein, the similarity is cosine similarity, with a value range of 0-1; when the similarity is greater than a first similarity threshold, a conditional predicate is added for differentiation, with the first similarity threshold ranging from 0.6 to 0.8; when the similarity is less than a second similarity threshold, the rule with higher confidence is adopted, with the second similarity threshold ranging from 0.2 to 0.

4.

6. An AI-based multi-factory server intelligent production scheduling system, used to execute the method according to any one of claims 1 to 5, characterized in that, include: Multiple factory edge modules and one central server module; The factory edge module includes: An encoding unit is used to compress the original production state into a low-dimensional representation vector through a preference-aware information bottleneck encoder. The encoding unit includes an adjustable compression coefficient and a preference alignment coefficient. Policy units are used to generate actions via a local policy network; A rule extraction unit is used to extract symbolic rules from the strategy unit to form a local rule set; The first communication unit is used to send the low-dimensional representation vector and the local rule set to the central server module, and to receive the scheduling decision and compression coefficient issued by the central server module. The central server module includes: The preference reasoning unit is used to receive the low-dimensional representation vectors and historical scheduling trajectories of all factories, and output the implicit preference vector, implicit constraints and uncertainty metric of preference reasoning for each factory. The uncertainty metric is a normalized value with a range of 0-1. A bitrate control unit, connected to the preference inference unit, is used to calculate the compression coefficient of each factory based on the uncertainty metric value, wherein the compression coefficient ranges from 0.01 to 1.

00. The rule generalization and resolution unit, connected to the preference reasoning unit, is used to receive the implicit preference vector and the local rule set of all factories, perform generalization processing and conflict resolution on the local rule set, and output the global rule base. The scheduling decision unit includes a rule matching subunit and a neural network subunit. The rule matching subunit receives the global rule base and the low-dimensional representation vectors of each factory, and outputs the matching rule and confidence level. The neural network subunit is activated when there is no matching rule or the confidence level is lower than a preset threshold, and outputs a scheduling decision. The preset threshold ranges from 0.5 to 0.

7. The rule evolution unit, connected to the scheduling decision unit, is used to receive the scheduling decision and execution feedback data output by the neural network subunit and output new rules; The second communication unit is paired with the first communication unit to exchange data. It is used to send scheduling decisions and compression coefficients to the edge modules of each factory, and to receive the low-dimensional representation vector and the local rule set.

7. The AI-based multi-factory server intelligent production scheduling system as described in claim 6, characterized in that, The central server module also includes a constraint verification unit, which is connected to the scheduling decision unit and the preference reasoning unit respectively. After the scheduling decision unit outputs a scheduling decision, it verifies whether the scheduling decision violates the implicit constraint conditions of any factory. If it does, it triggers the scheduling decision unit to regenerate the scheduling decision.

8. The AI-based multi-factory server intelligent production scheduling system as described in claim 6, characterized in that, The rule evolution unit also includes a confirmation interface, which is used to visually display the logical structure of the new rule and receive confirmation, modification or rejection instructions.

9. The AI-based multi-factory server intelligent production scheduling system as described in claim 6, characterized in that, The preference inference unit also periodically sends the implicit preference vector and uncertainty metric value obtained from the current inference to each factory edge module through the second communication unit; the encoding unit receives the implicit preference vector and uncertainty metric value and adjusts its preference alignment coefficient accordingly. The preference alignment coefficient ranges from 0.1 to 1.

0.

10. The AI-based multi-factory server intelligent production scheduling system as described in claim 6, characterized in that, The central server module also includes a data storage unit for persistently storing the historical scheduling trajectories, low-dimensional representation vectors, global rule base, and rule evolution logs of all factories; the storage retention period set for the data storage unit is 30-365 days.

Images