A global forward-looking intelligent scheduling method and system for a complex dynamic workshop

Abstract

The application discloses a kind of global prospective intelligent scheduling methods and systems for complex dynamic workshop, the method includes: the original data of workshop collected is handled according to industrial grade data preprocessing rule, constructs the real-time state feature vector of current process;According to the process route information of to-be-scheduled process, extract multi-step future process path information and encode as structured feature vector;Real-time state feature vector and structured feature vector are weighted and fused, and generate fusion state representation;The preset intelligent inference model is inputted into fusion state representation, and output structured scheduling instruction;Structured scheduling instruction is issued to workshop execution layer, and continuously receives execution feedback information, and based on actual execution state and dynamic disturbance event triggers global prospective dynamic scheduling.The application encodes multi-step future process path information as structured feature vector and is weighted and fused with real-time state, so that scheduling decision can be explicitly predicted long-term resource competition and process constraint at the moment.

Description

Technical Field

[0001] This application relates to the field of workshop scheduling technology, and more specifically, to a global forward-looking intelligent scheduling method and system for complex dynamic workshops. Background Technology

[0002] Shop floor scheduling is a core component of intelligent manufacturing systems. Its task is to rationally allocate processing equipment, arrange the sequence of processes, and determine start and end times, while meeting process routes, resource constraints, and production targets, in order to maximize production efficiency, minimize costs, and improve on-time delivery rates. As the manufacturing industry transforms towards flexibility and personalization, multi-variety, small-batch production models are becoming increasingly prevalent. Shop floor operations are exhibiting highly dynamic and uncertain characteristics, with frequent disruptions such as equipment failures, emergency order insertions, and material delays, posing significant challenges to the real-time response and global optimization capabilities of scheduling methods.

[0003] Currently, mainstream workshop scheduling methods can be mainly divided into three categories: First, heuristic rule-based methods, such as earliest delivery date priority and shortest processing time priority, are simple to calculate and have a fast response, but their decision-making perspective is limited to the current process and lacks overall consideration of the global state and future evolution. They are prone to getting trapped in local optima in complex dynamic environments. Second, mathematical programming-based methods, such as mixed integer linear programming and constrained programming, can find theoretically optimal solutions in static scenarios, but their computational complexity increases exponentially with the problem size, and once the model is established, it is difficult to adapt to dynamic changes and cannot meet the real-time scheduling requirements. Third, metaheuristic algorithm-based methods, such as genetic algorithms and simulated annealing, achieve a certain balance between solution quality and efficiency, but their performance is highly dependent on parameter tuning and they mostly adopt offline optimization paradigms, making it difficult to effectively represent the chain effect of the current choice on the future multi-step production state during the decision-making process.

[0004] The common drawback of the aforementioned methods is that their scheduling decisions rely on a single information dimension and a short-sighted temporal perspective. They can only make reactive decisions based on the local state at the current moment, and cannot explicitly predict resource competition, bottleneck shifts, and long-term dependencies in future multi-step paths during the process allocation stage. Existing solutions lack a technical means to structurally express future process path constraints and deeply integrate them with real-time states. As a result, the scheduling system lacks true foresight and global perspective. When faced with dynamic disturbances, it can often only respond passively and make local repairs, making it difficult to proactively avoid potential production conflicts and efficiency losses from a global perspective. Summary of the Invention

[0005] To address at least one deficiency or improvement need in the prior art, this invention provides a global forward-looking intelligent scheduling method and system for complex dynamic workshops. This addresses the problems in the prior art where scheduling decisions rely on a single information dimension and a short-sighted time perspective, making reactive decisions based only on the local state at the current moment. When faced with dynamic disturbances, these systems often only offer passive responses and local repairs, making it difficult to proactively avoid potential production conflicts and efficiency losses from a global perspective.

[0006] To achieve the above objectives, according to a first aspect of the present invention, a global forward-looking intelligent scheduling method for complex dynamic workshops is provided, comprising: The collected raw workshop data is processed according to industrial-grade data preprocessing rules to construct the real-time status feature vector of the current process. Based on the process route information of the process to be scheduled, extract the multi-step future process path information and encode it into a structured feature vector; The real-time state feature vector and the structured feature vector are weighted and fused to generate a fused state representation; The fused state representation is input into a preset intelligent reasoning model, and structured scheduling instructions are output. Structured scheduling instructions are sent to the workshop execution layer, and execution feedback information is continuously received. Global forward-looking dynamic scheduling is triggered based on the actual execution status and dynamic disturbance events.

[0007] In one possible implementation, the collected raw workshop data is processed according to industrial-grade data preprocessing rules, and also includes: The system determines whether the deviation between the original workshop data and the historical average exceeds a preset fluctuation threshold and smooths out noisy data. Abnormal data is marked and removed according to preset process rules, and historical statistical values ​​are used to fill in the missing data after removal; The filled data is transformed into a unified numerical range using a linear mapping method to generate a standardized global snapshot of the workshop status.

[0008] In one possible implementation, constructing the real-time state feature vector of the current process also includes: Obtain the status information of each selectable processing equipment for the workpiece corresponding to the current process from the global status snapshot of the workshop, the number of waiting processing tasks in the equipment queue, and the estimated idle time of the equipment; Obtain the current inventory level of materials required for the current process from the global snapshot of the shop floor status; The status information of the selectable processing equipment, the number of waiting processing tasks, the estimated idle time, and the current inventory level are concatenated and vectorized to form a real-time status feature vector.

[0009] In one possible implementation, based on the process route information of the process to be scheduled, multi-step future process path information is extracted and encoded into a structured feature vector, and the implementation further includes: Extract all subsequent processes after the process to be scheduled based on the process route information to form a future process sequence; For each subsequent process in the future process sequence, extract the process identifier, the set of optional processing equipment, the standard processing time, the time offset relative to the current process, and the demand intensity vector of various production resources as feature information of the subsequent process. A two-dimensional future path information matrix is ​​constructed, with each subsequent process in the future process sequence as a row and the feature information corresponding to each subsequent process as a column; A pre-trained neural network encoder maps a two-dimensional future path information matrix into a structured feature vector.

[0010] In one possible implementation, the real-time state feature vector and the structured feature vector are weighted and fused to generate a fused state representation, which also includes: The real-time state feature vector and structured features are input into an attention-based feature fusion network; Dynamic attention weights are computed for each feature dimension in the real-time state feature vector and the structured feature vector through a feature fusion network; Based on dynamic attention weights, the real-time state feature vector and the structured feature vector are weighted and summed, and then subjected to nonlinear transformation to output the fused state representation.

[0011] In one possible implementation, the pre-defined intelligent reasoning model is a large language model that has been fine-tuned with in-depth knowledge of the workshop scheduling domain. The input layer of a large language model receives the fusion state representation, the weighted dependency matrix output by the dependency quantization unit, and the device load level characteristics output by the temporal prediction unit. The weighted dependency matrix is ​​used to provide structural constraint information between processes caused by resource competition to large language models; Equipment load level features are used to provide information on the load stress of critical equipment in future time periods to large language models.

[0012] In one possible implementation, structured scheduling instructions are issued to the shop floor execution layer, and execution feedback information is continuously received. Global forward-looking dynamic scheduling is triggered based on the actual execution status and dynamic disturbance events. This also includes: It continuously receives real-time production data from the workshop execution layer to monitor for any type of disturbance event, such as equipment failure, emergency order insertion, or material shortage, or whether the preset periodic scheduling and inspection time point has been reached. When a disturbance event occurs or a periodic scheduling check point arrives, calculate the deviation between the earliest start time of each process and the baseline planned time, and estimate the impact of the disturbance on the total completion time or order delivery date. The impact level is compared with a preset impact threshold. If the impact level exceeds the preset impact threshold, the global forward-looking dynamic scheduling process is triggered to update the scheduling instructions. If the impact does not exceed the preset impact threshold, the planned start and end times of the affected processes will be adjusted based on the original scheduling plan.

[0013] According to a second aspect of the present invention, a global forward-looking intelligent scheduling system for complex dynamic workshops is also provided, comprising: The preprocessing module is configured to process the collected raw workshop data according to industrial-grade data preprocessing rules to construct the real-time status feature vector of the current process. The encoding module is configured to extract multi-step future process path information and encode it into a structured feature vector based on the process route information of the process to be scheduled. The weighting module is configured to perform weighted fusion of real-time state feature vectors and structured feature vectors to generate a fused state representation; The instruction generation module is configured to input the fused state representation into a preset intelligent reasoning model and output structured scheduling instructions. The feedback adjustment module is configured to send structured scheduling instructions to the workshop execution layer and continuously receive execution feedback information, triggering global forward-looking dynamic scheduling based on the actual execution status and dynamic disturbance events.

[0014] According to a third aspect of the present invention, a global forward-looking intelligent scheduling device for complex dynamic workshops is also provided, comprising at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program, and when the computer program is executed by the processing unit, the processing unit performs the steps of any of the above-described global forward-looking intelligent scheduling methods for complex dynamic workshops.

[0015] According to a fourth aspect of the present invention, a storage medium is also provided, which stores a computer program executable by a global forward-looking intelligent scheduling device for complex dynamic workshops. When the computer program is run on the global forward-looking intelligent scheduling device for complex dynamic workshops, the global forward-looking intelligent scheduling device for complex dynamic workshops executes the steps of any of the above-described global forward-looking intelligent scheduling methods for complex dynamic workshops.

[0016] In summary, compared with the prior art, the above-described technical solutions conceived by this invention can achieve the following beneficial effects: This invention provides a global forward-looking intelligent scheduling method for complex dynamic workshops. It extracts multi-step future process path information and encodes it into structured feature vectors. It obtains information on multiple unexecuted subsequent processes after the scheduled process from the process route and organizes this information into structured feature vectors. This expands the information upon which scheduling decisions rely from the current single process to a sequence of future processes, transforming long-term constraints in the time dimension into currently computable feature expressions. The real-time state feature vector and the structured feature vector are weighted and fused, jointly representing the real-time state and future path information. This allows the model to simultaneously possess a complete information view of both the present and future dimensions within the same decision space. The fused state representation, including future multi-step path constraints, is directly input into the intelligent inference model. The model completes the entire decision-making process from global state perception to scheduling instruction generation end-to-end. Within a complete decision space encompassing the current state and future constraints, it autonomously learns scheduling strategies, proactively identifies potential resource competition, path congestion, and bottleneck transfer risks, and makes avoidance decisions in advance at the moment of process allocation. When a disturbance occurs, the impact of the disturbance is assessed based on a global forward-looking perspective, and a decision is made on whether to re-execute the complete scheduling process. This makes dynamic scheduling no longer a reactive, fragmented repair, but a controlled adjustment that always revolves around the global optimization goal, thus achieving the continuous maintenance and autonomous repair of the global production order in a dynamic environment. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart illustrating an embodiment of the global forward-looking intelligent scheduling method for complex dynamic workshops provided by the present invention. Figure 2 A flowchart illustrating an embodiment of the encoding of future process path information provided by the present invention; Figure 3 A flowchart illustrating an embodiment of attention mechanism feature fusion provided by the present invention; Figure 4 A schematic diagram of an embodiment of the global forward-looking intelligent scheduling system for complex dynamic workshops provided by the present invention; Figure 5 This is a schematic diagram of the structure of a global forward-looking intelligent scheduling device for complex dynamic workshops provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention. Furthermore, the technical features involved in the various embodiments of this invention described below can be combined with each other as long as they do not conflict with each other.

[0020] The terms "first," "second," "third," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0021] This invention provides a global forward-looking intelligent scheduling method and system for complex dynamic workshops, which will be described below.

[0022] Please see Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the global forward-looking intelligent scheduling method for complex dynamic workshops provided by the present invention. In a specific embodiment of the present invention, a global forward-looking intelligent scheduling method for complex dynamic workshops is disclosed, including: S101. Process the collected raw workshop data according to industrial-grade data preprocessing rules to construct the real-time status feature vector of the current process. S102. Based on the process route information of the process to be scheduled, extract the multi-step future process path information and encode it into a structured feature vector; S103. Weighted fusion of real-time state feature vector and structured feature vector to generate fused state representation; S104. Input the fusion state representation into the preset intelligent reasoning model and output the structured scheduling instruction; S105. The structured scheduling instructions are sent to the workshop execution layer and the execution feedback information is continuously received. Based on the actual execution status and dynamic disturbance events, global forward-looking dynamic scheduling is triggered.

[0023] In the above embodiments, raw production data of the entire workshop is collected in real time through manufacturing execution systems, equipment control systems or industrial sensor networks. The raw production data includes at least the real-time operating status codes of each processing equipment, the number of tasks waiting to be processed in the equipment queue, the expected time node for the equipment to enter the idle state, the processing progress of the workpiece corresponding to the current process, and the inventory level of the key materials required for the process.

[0024] Due to the complexity of industrial environments, raw data often contains transient fluctuations, sensor noise, communication packet loss, and abnormal values ​​that violate process rules. Directly using this data for scheduling decisions would severely impact the stability and reliability of the model. Therefore, it is necessary to execute an industrial-grade data preprocessing rule set to generate an accurate, clean, and aligned snapshot of the shop floor's global state. Based on this, for the current process to be scheduled, the status code, queue length, and estimated idle time of each selectable processing equipment are extracted from this state snapshot, along with the real-time inventory level of the materials required for that process. These are then concatenated into a high-dimensional numerical vector, i.e., the real-time state feature vector, according to predefined encoding rules and order.

[0025] To achieve truly proactive scheduling, the process route database is first queried to identify all unexecuted subsequent processes following the workpiece to be scheduled, forming an ordered sequence of future processes. For each subsequent process in this sequence, the following five types of descriptive information are extracted: 1) A unique identifier for the process, used to distinguish different process entities; 2) The set of processing equipment that can be selected for the process under process constraints, represented in the form of multi-hot encoding or embedded vectors; 3) The estimated standard processing time for the process on each selectable equipment, which can be derived from process document presets or historical statistical corrections; 4) The time offset or sequence position index of the process relative to the current process, used to characterize the sequential dependency relationship between processes; 5) The demand intensity vector of the process for various production resources, including but not limited to equipment type, special tooling, operator qualifications, and material categories.

[0026] The five types of information from each subsequent process are concatenated into a row vector in a fixed order. All row vectors from subsequent processes are then stacked from top to bottom according to their order in the process route, forming a two-dimensional future path information matrix. The number of rows in the matrix equals the number of subsequent processes, and the number of columns equals the total dimension after encoding the five types of information. A pre-trained neural network encoder is used to extract features and compress the dimension of this two-dimensional matrix. The encoder can employ a recurrent neural network to capture long-range dependencies in the process sequence, or a one-dimensional convolutional neural network to extract pattern features within a local sliding window. After forward propagation computation by the encoder, the original variable-length two-dimensional matrix is ​​mapped into a dense numerical vector of fixed dimension; this vector is the structured feature vector.

[0027] A feature fusion network based on an attention mechanism is introduced, taking real-time state feature vectors and structured feature vectors as inputs. A linear transformation maps both to the same semantic space, calculating the correlation scores between dimensions within each feature vector and across vectors. The attention mechanism uses one feature vector as the query and the other as the key and value, calculating the weight distribution through dot product or additive attention. This reflects which dimensions in the real-time state are more strongly correlated with which dimensions in the future path under the current scheduling decision context. For example, when the structured feature vector implicitly encodes that a key future process is highly dependent on a particular device, the attention mechanism automatically increases the weights of dimensions such as device status, queue length, and idle time in the real-time state feature vector; conversely, if the demand for a certain type of resource in the future path is low, the attention to the corresponding real-time state features will be suppressed.

[0028] After attention weighting and residual connection, a new vector that integrates both types of information is output. Then, feature transformation is performed through one or more fully connected networks and nonlinear activation functions to finally generate the fused state representation. This representation includes real-time production information such as current equipment status and material inventory, as well as process constraints, resource competition risks, and time pressures for future multi-step processes, achieving a deep integration of the present and the future in the feature space.

[0029] Unlike general language models, intelligent reasoning models undergo specialized domain adaptation during the training phase. First, through supervised fine-tuning, the basic mapping ability from state representation to scheduling actions is learned on a large-scale workshop state-optimal scheduling instruction labeled dataset. Then, through reinforcement learning fine-tuning, the model strategy is optimized in interaction with the simulation environment, using a weighted combination of multiple production performance indicators such as equipment utilization, total completion time, order delay penalty, and total critical path deviation as reward signals. This allows the model to pursue the maximization of long-term cumulative rewards, thereby learning the global optimization decision-making ability for future multi-step processes.

[0030] The generated structured scheduling instructions are parsed by the instruction generation module and converted into executable instructions that conform to the communication protocol of the underlying equipment in the workshop. These instructions are then distributed to the corresponding processing equipment, material handling system, or operator terminals via industrial Ethernet or wireless communication networks. The execution layer organizes production activities according to the instructions. Simultaneously, the system continuously receives real-time feedback data from the workshop through the status awareness module. This feedback data includes equipment status change events, process start and completion times, deviations between actual and planned processing times, material consumption, and various abnormal alarm signals.

[0031] Building upon this foundation, the present invention also establishes a lightweight and global collaborative dynamic scheduling triggering mechanism, continuously monitoring two types of triggering conditions: first, whether a preset type of dynamic disturbance event occurs, including but not limited to sudden equipment failure, high-priority emergency order insertion, shortage of critical materials, and temporary changes to the process route; and second, whether a preset periodic scheduling check time point has been reached. When either triggering condition is met, a lightweight evaluation process is initiated. Based on the current actual execution state, forward propagation calculations are performed only on the affected nodes on the critical production path to quickly estimate the deviation value of the disturbance event's impact on key performance indicators such as total completion time and order delivery date, and this deviation value is compared with a preset tolerance threshold.

[0032] If the deviation value does not exceed the threshold, the disturbance is considered controllable. Only the planned start and end times of relevant processes are adjusted sequentially based on the original scheduling plan, without triggering a complete global rescheduling process, thus achieving rapid response with minimal computational overhead. If the deviation value exceeds the threshold, the disturbance is considered to have a substantial impact on the overall production order, triggering a complete global forward-looking dynamic scheduling. This involves re-executing all steps S101 to S104, generating a new scheduling plan adapted to the new environmental constraints, starting from the latest global workshop state. This allows the scheduling system to maintain plan stability and computational economy in the face of minor disturbances, while promptly activating global optimization capabilities in the face of major disturbances, achieving an adaptive balance between response speed and decision quality.

[0033] Compared with existing technologies, this embodiment provides a global forward-looking intelligent scheduling method for complex dynamic workshops. By extracting multi-step future process path information and encoding it into structured feature vectors, it obtains information on multiple subsequent processes that have not yet been executed after the process to be scheduled from the process route and organizes them into structured feature vectors. This expands the information upon which scheduling decisions rely from the current single process to a sequence of future multi-step processes, transforming long-term constraints in the time dimension into currently computable feature expressions. The real-time state feature vector and the structured feature vector are weighted and fused, jointly representing the real-time state and future path information. This allows the model to simultaneously possess a complete information view of both the present and future dimensions within the same decision space. The fused state representation, including future multi-step path constraints, is directly input into the intelligent inference model. The model completes the entire decision-making process from global state perception to scheduling instruction generation end-to-end. Within a complete decision space containing the current state and future constraints, it autonomously learns scheduling strategies, proactively identifies potential resource competition, path congestion, and bottleneck transfer risks, and makes avoidance decisions in advance at the moment of process allocation. When a disturbance occurs, the impact of the disturbance is assessed based on a global forward-looking perspective, and a decision is made on whether to re-execute the complete scheduling process. This makes dynamic scheduling no longer a reactive, fragmented repair, but a controlled adjustment that always revolves around the global optimization goal, thus achieving the continuous maintenance and autonomous repair of the global production order in a dynamic environment.

[0034] In some embodiments of the present invention, the collected raw workshop data is processed according to industrial-grade data preprocessing rules, and the process further includes: The system determines whether the deviation between the original workshop data and the historical average exceeds a preset fluctuation threshold and smooths out noisy data. Abnormal data is marked and removed according to preset process rules, and historical statistical values ​​are used to fill in the missing data after removal; The filled data is transformed into a unified numerical range using a linear mapping method to generate a standardized global snapshot of the workshop status.

[0035] In some embodiments of the present invention, constructing the real-time state feature vector of the current process further includes: Obtain the status information of each selectable processing equipment for the workpiece corresponding to the current process from the global status snapshot of the workshop, the number of waiting processing tasks in the equipment queue, and the estimated idle time of the equipment; Obtain the current inventory level of materials required for the current process from the global snapshot of the shop floor status; The status information of the selectable processing equipment, the number of waiting processing tasks, the estimated idle time, and the current inventory level are concatenated and vectorized to form a real-time status feature vector.

[0036] Please see Figure 2 , Figure 2 This is a flowchart illustrating an embodiment of the encoding of future process path information provided by the present invention. In some embodiments of the present invention, based on the process route information of the process to be scheduled, multi-step future process path information is extracted and encoded into a structured feature vector, and the method further includes: Extract all subsequent processes after the process to be scheduled based on the process route information to form a future process sequence; For each subsequent process in the future process sequence, extract the process identifier, the set of optional processing equipment, the standard processing time, the time offset relative to the current process, and the demand intensity vector of various production resources as feature information of the subsequent process. A two-dimensional future path information matrix is ​​constructed, with each subsequent process in the future process sequence as a row and the feature information corresponding to each subsequent process as a column; A pre-trained neural network encoder maps a two-dimensional future path information matrix into a structured feature vector.

[0037] Please see Figure 3 , Figure 3 This is a flowchart illustrating an embodiment of the attention mechanism feature fusion provided by the present invention. In some embodiments of the present invention, the real-time state feature vector and the structured feature vector are weighted and fused to generate a fused state representation, and the method further includes: The real-time state feature vector and structured features are input into an attention-based feature fusion network; Dynamic attention weights are computed for each feature dimension in the real-time state feature vector and the structured feature vector through a feature fusion network; Based on dynamic attention weights, the real-time state feature vector and the structured feature vector are weighted and summed, and then subjected to nonlinear transformation to output the fused state representation.

[0038] In some embodiments of the present invention, the preset intelligent reasoning model is a large language model that has been fine-tuned with depth of knowledge in the workshop scheduling domain; The input layer of a large language model receives the fusion state representation, the weighted dependency matrix output by the dependency quantization unit, and the device load level characteristics output by the temporal prediction unit. The weighted dependency matrix is ​​used to provide structural constraint information between processes caused by resource competition to large language models; Equipment load level features are used to provide information on the load stress of critical equipment in future time periods to large language models.

[0039] In some embodiments of the present invention, structured scheduling instructions are issued to the shop floor execution layer, and execution feedback information is continuously received. Global forward-looking dynamic scheduling is triggered based on the actual execution status and dynamic disturbance events. The method further includes: It continuously receives real-time production data from the workshop execution layer to monitor for any type of disturbance event, such as equipment failure, emergency order insertion, or material shortage, or whether the preset periodic scheduling and inspection time point has been reached. When a disturbance event occurs or a periodic scheduling check point arrives, calculate the deviation between the earliest start time of each process and the baseline planned time, and estimate the impact of the disturbance on the total completion time or order delivery date. The impact level is compared with a preset impact threshold. If the impact level exceeds the preset impact threshold, the global forward-looking dynamic scheduling process is triggered to update the scheduling instructions. If the impact does not exceed the preset impact threshold, the planned start and end times of the affected processes will be adjusted based on the original scheduling plan.

[0040] The entire process begins with real-time, precise sensing of the overall status of the workshop. The system, through its built-in data sensing and integration module, communicates continuously with the Manufacturing Execution System (MES), the Supervisory Control System (SCADA), and various industrial sensor networks distributed throughout the workshop via industrial Ethernet protocols such as Profinet and OPC-UA, or wireless networks such as 5G and Wi-Fi.

[0041] All this raw data is processed using an industrial-grade data preprocessing rule set to ensure high accuracy and consistency. The industrial-grade data preprocessing rule set includes: Transient noise filtering: Set a fluctuation threshold (e.g.) If the data collected in a single instance is different from the previous one... If the mean deviation exceeds the threshold, the previous method is used. Substituting the mean avoids noise introduced by instantaneous fluctuations.

[0042] Regularized outlier hard removal and data completion: Preset outlier thresholds (e.g., device status codes exceeding the valid set, processing time is...). Or exceeding the historical maximum value Abnormal data is directly marked and removed, and then filled in through a data completion mechanism (such as historical averages or comparison data of similar devices).

[0043] Fixed interval standardization: Define the physical range of values ​​according to process requirements, and use linear mapping to uniformly map the data to intervals.

[0044] Subsequently, these data are aggregated to form a precise, dynamically updated snapshot of the overall status of the workshop.

[0045] Next, the method enters its crucial feature engineering phase. To achieve forward-looking decision-making, two closely related feature vectors are innovatively constructed. First, based on a global snapshot of the shop floor state, a real-time state feature vector is generated for each process to be scheduled, denoted as... This vector is a high-dimensional numerical array, and its structure can be exemplified as follows:

[0046] in, Indicates optional machine Status codes (e.g., 0 represents idle, 1 represents processing). Representative machine The current number of queued tasks, Estimate their free time. Represents the current process Inventory levels of required materials.

[0047] Furthermore, in order to enable the decision-making model to have a forward-looking understanding of the subsequent production evolution when making decisions in the current process, the system extracts the sequence information of all subsequent unfinished processes from the process route database for the process to be scheduled, and encodes them into a structured future process path information matrix.

[0048] The future process path information matrix is ​​structured with process sequences as rows, each row corresponding to a subsequent process. Its column vectors describe the key constraints of that process within the production system. Specifically, each row includes at least the following information dimensions: (1) Process identification features, used to uniquely identify the position of the process in the process route; (2) Optional processing resource coding features, which are used to characterize the set of machines or equipment that can be selected under process constraints for this process. These features can be represented in the form of multi-hot coding, index mapping or embedding vectors. (3) Estimated processing time characteristics, used to characterize the typical processing time of the process on different optional resources. The values ​​can be derived from process documents, historical statistics or online correction results; (4) Time offset or sequence constraint features are used to describe the relative position of the process with respect to the current process in the process route, so as to reflect the sequential dependence between processes; (5) Resource demand vector features are used to characterize the demand for key production resources in the process, including but not limited to resource dimensions such as equipment type, tooling, materials or manpower.

[0049] The resource demand vector and the optional processing resource encoding together constitute a constraint description of future processes in the resource dimension, enabling the model to explicitly perceive the potential competitive relationships and conflict risks of different future processes on specific resources. This type of structural information cannot be fully expressed by a single time series feature and is a key foundation for realizing global forward-looking scheduling reasoning.

[0050] To process this variable-length and heterogeneous future information matrix, the system employs a neural network model, denoted as . This is compressed into a fixed-dimensional structured feature vector. ;in These are the model parameters.

[0051] The purpose of this encoding process is not for conventional classification or prediction, but to generate a holistic latent representation of future multi-step constraints. This representation is specifically designed for dynamic weighting by subsequent attention mechanisms and ultimately serves high-level logical reasoning in large language models. This vector It can capture long-range dependencies throughout the entire future process sequence.

[0052] To enable the model to understand both the present and the future simultaneously, the system employs a feature fusion network based on an attention mechanism. The network receives real-time state feature vectors. and structured feature vectors As input, it dynamically evaluates the importance of each feature through its inherent attention weight allocation mechanism and outputs a fused state representation.

[0053] For example, when The study reveals that when a key process becomes highly dependent on a particular machine, the attention mechanism will automatically improve. The weights of the current state characteristics of this machine.

[0054] Subsequently, the fused state representation is input into the intelligent reasoning model for scheduling decision reasoning. The intelligent reasoning model is a large-scale language model that has undergone deep fine-tuning for the workshop scheduling domain. During the reasoning process, it invokes a dependency quantization unit and a time-series prediction unit to generate auxiliary decision features.

[0055] The dependency quantification unit constructs a process-resource dependency matrix based on process route information and resource constraint information, and calculates a weighted dependency representation based on the dependency strength of each process on the corresponding resource, which is used to represent the constraint relationship of different processes in the resource dimension; the time series prediction unit performs statistical analysis on equipment load based on historical production data, and corrects the prediction results by combining real-time disturbance events such as emergency order insertion and equipment failure, and outputs load level characteristics representing the future load status of the equipment.

[0056] The weighted dependency representation, load level feature, and fused state representation are combined and encoded to form a multi-dimensional state input for scheduling decision-making. This multi-dimensional state input is then provided to the intelligent inference model, which generates the corresponding scheduling decision result based on the multi-dimensional state input.

[0057] The scheduling decision results are used to determine the allocation relationship between processes and processing resources and their planned execution time. When multiple candidate scheduling paths exist, the intelligent reasoning model comprehensively evaluates the impact of different scheduling paths on resource consumption and load distribution based on the process-resource dependency representation and equipment load level characteristics, thereby determining a scheduling scheme that meets global constraints.

[0058] To enhance the decision-making capability of the intelligent reasoning model in complex and dynamic workshop environments, the model employs a multi-stage training approach for optimization. The first stage uses supervised fine-tuning to enable the model to learn the mapping relationship between workshop states and scheduling instructions. The second stage employs a proximal policy optimization algorithm or a group relative policy optimization algorithm for reinforcement learning training. During reinforcement learning training, a state input is constructed containing the current scheduling process and at least two subsequent processes. Based on the scheduling results, the total completion time, resource utilization, and critical path deviation are evaluated as the basis for updating the model parameters.

[0059] After training, the intelligent inference model outputs structured scheduling instructions, which include at least a workpiece identifier, a process identifier, a target processing equipment, and corresponding planned start and end times, for subsequent scheduling execution.

[0060] After deep inference, the large model ultimately outputs a series of structured scheduling instructions, which typically follow JSON format, as shown in the following example: { "command_id": "CMD-20250929-001", "work_order_id": "J1","operation_id": "OP2", "target_machine_id": "M5", "planned_start_time":"2025-09-29T17:10:00Z","planned_end_time":"2025-09-29T17:50:00Z"}.

[0061] Finally, this invention constructs an adaptive closed-loop control and dynamic scheduling mechanism. The system continuously receives execution feedback and triggers dynamic scheduling using a hybrid strategy combining event-driven and periodic-driven approaches. When a disturbance event occurs, the system first initiates a lightweight evaluation process to quickly estimate the impact of the disturbance on the overall plan without invoking the computationally expensive full intelligent inference model's reasoning process.

[0062] This lightweight assessment process is based on critical path deviation calculation: it extracts the critical path from the MES, sets a baseline completion time for each critical process, and when a disturbance event occurs, it quickly estimates the impact of the disturbance on key performance indicators (KPIs) such as total completion time and delivery date through critical path deviation calculation. This calculation only updates the earliest start time of subsequent task nodes on the affected critical path, without changing the already determined process sequence or resource allocation.

[0063] Only when the critical path delay or core equipment load deviation exceeds a preset threshold. (For example, if the total completion time is extended by more than 5%), the system will determine that the disturbance has a significant impact and trigger the full global dynamic scheduling process of S4. If the estimated degree of deterioration is within an acceptable range, the system will only perform local adjustments, that is, directly adopt the new schedule.

[0064] Only when the KPI deterioration level estimated by this process exceeds a preset threshold. (For example, if the total completion time is extended by more than 5%), the system will determine that the disturbance has a significant impact and initiate a full global forward-looking dynamic scheduling. If the estimated degree of deterioration is within an acceptable range, the system will only perform local adjustments, that is, directly adopt the new schedule calculated by forward propagation and update the planned start and end times of the affected processes. This design achieves an intelligent balance between response speed and computational overhead, ensuring a rapid and economical response to minor disturbances while retaining the ability to perform global optimization in the event of significant changes.

[0065] To illustrate the invention more clearly, the following will use a specific production scenario to explain the operational details of the entire process.

[0066] Suppose that at 5:00 PM, the workshop faces a complex scheduling decision. At this time, workpiece J1 is awaiting its operation OP2, with machine options M2 and M5. Real-time status data from the system shows that machine M2 is currently idle, while machine M5 is processing another workpiece J2, expected to be completed in 10 minutes (i.e., 5:10 PM). Simultaneously, the workshop has just received a high-priority urgent workpiece J3, whose first operation OP31 also requires machine M2. Furthermore, according to the process routing database, a key constraint is that the subsequent critical operation OP4 of workpiece J1 must be completed on machine M2 before 6:00 PM.

[0067] Faced with this situation, the method of this invention begins to perform forward scheduling. First, the system constructs features for the J1-OP2 operations to be scheduled. Based on the current state, its real-time state feature vector... Will contain Information such as these. More importantly, the system will construct a matrix of its future path information. This matrix explicitly encodes subsequent constraints, such as [[ [M2], 40, 60,...], ...]. This matrix is ​​encoded by an encoder. The vector is compressed to form a structured feature vector that emphasizes "M2's key mission in the future". .

[0068] Subsequently, in the feature fusion and inference stage, the real-time state vector and future feature vectors Sent into the converged network .because There are strong constraints on M2, and the attention mechanism significantly increases the weight of real-time state features related to M2. The fused representation The information is passed to the intelligent reasoning model. At this point, the model performs deep reasoning: while immediately allocating the idle M2 to J1-OP2 seems like an efficient "greedy" decision, the model, incorporating future information, can foresee the negative chain reaction of this action. The model analyzes that J1's subsequent critical process OP4 urgently needs M2 within an hour, while the high-priority urgent task J3 is also waiting for M2. If M2 is used now, it will not only delay J3 but may also prevent M2 from preparing for J1's subsequent critical process OP4 in time, thus causing a more serious global bottleneck. Therefore, after weighing the pros and cons, the model chooses a globally better solution: let J1 wait 10 minutes and use the soon-to-be-idle M5 instead. Although this decision sacrifices J1's immediate benefit, it frees up a valuable processing window for the urgent J3 and reserves sufficient time for J1's subsequent critical processes. This perfectly aligns with the decision-making strategy learned during the reinforcement learning (PPO / GRPO) fine-tuning phase, which aims to maximize long-term cumulative benefits. Based on this reasoning, the model ultimately generated and issued structured JSON scheduling instructions for J1 and J3.

[0069] The dynamic scheduling capability of this invention can be demonstrated when a disturbance event occurs. For example, at a certain moment, machine M5 suddenly reports a mechanical failure. This event will immediately activate the dynamic scheduling mechanism. The system first discovers through a lightweight assessment that if the original plan is maintained, the total completion time is expected to increase by more than 50%, far exceeding the preset 5% threshold. The system determines that a global dynamic scheduling is needed, and immediately re-executes the complete process from feature construction to inference for J1 and other affected artifacts with the updated state (M5 becomes a fault code). Under the new constraints, large language models will be based on In the new scenario where only M2 is available, the optimal processing order of J1 and J3 on M2 is recalculated, and a completely new scheduling scheme adapted to the failure event is output.

[0070] In another scenario, if the disturbance is minor, the system's response differs. Suppose M5 does not malfunction, but at some point the system receives feedback that the actual time taken for process J3-OP31 on M2 is 3 minutes longer than planned. This deviation would also trigger a dynamic scheduling mechanism, but a lightweight evaluation simulation shows that the delay only increases the total completion time by 1.2%, below the 5% threshold. Therefore, the system determines that global dynamic scheduling is unnecessary and instead initiates a local adjustment procedure, only postponing the planned start time of all subsequent tasks on M2 by 3 minutes. This approach avoids unnecessary global calculations, responds quickly to minor changes with minimal computational overhead, ensures the overall stability of the production plan, and demonstrates the system's intelligent balance between responsiveness and economy.

[0071] To better implement the globally forward-looking intelligent scheduling method for complex dynamic workshops in this invention embodiment, based on the globally forward-looking intelligent scheduling method for complex dynamic workshops, please refer to the corresponding... Figure 4 , Figure 4 This is a schematic diagram of an embodiment of the global forward-looking intelligent scheduling system for complex dynamic workshops provided by the present invention. The embodiment of the present invention provides a global forward-looking intelligent scheduling system 400 for complex dynamic workshops, comprising: The preprocessing module 410 is configured to process the collected raw workshop data according to industrial-grade data preprocessing rules to construct the real-time status feature vector of the current process. The encoding module 420 is configured to extract multi-step future process path information and encode it into a structured feature vector based on the process route information of the process to be scheduled. The weighting module 430 is configured to perform weighted fusion of the real-time state feature vector and the structured feature vector to generate a fused state representation; The instruction generation module 440 is configured to input the fused state representation into a preset intelligent reasoning model and output structured scheduling instructions. The feedback adjustment module 450 is configured to send structured scheduling instructions to the workshop execution layer and continuously receive execution feedback information, triggering global forward-looking dynamic scheduling based on the actual execution status and dynamic disturbance events.

[0072] It should be noted that the system 400 provided in the above embodiments can implement the technical solutions described in the above method embodiments. The specific implementation principles of the above modules or units can be found in the corresponding content in the above method embodiments, and will not be repeated here.

[0073] Please see Figure 5 , Figure 5 This is a schematic diagram of the structure of a globally forward-looking intelligent scheduling device for complex dynamic workshops provided in an embodiment of the present invention. Based on the above-described globally forward-looking intelligent scheduling method for complex dynamic workshops, the present invention also provides a globally forward-looking intelligent scheduling device for complex dynamic workshops. This device can be a computing device such as a mobile terminal, desktop computer, laptop, handheld computer, or server. The globally forward-looking intelligent scheduling device 500 for complex dynamic workshops includes a processor 510, a memory 520, and a display 530. Figure 5 Only a portion of the components of a globally forward-looking intelligent scheduling device for complex dynamic workshops are shown. However, it should be understood that implementation of all the components shown is not required, and more or fewer components may be implemented instead.

[0074] In some embodiments, the memory 520 may be an internal storage unit of the global forward-looking intelligent scheduling device 500 for complex dynamic workshops, such as a hard disk or memory of the global forward-looking intelligent scheduling device 500 for complex dynamic workshops. In other embodiments, the memory 520 may be an external storage device of the global forward-looking intelligent scheduling device 500 for complex dynamic workshops, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the global forward-looking intelligent scheduling device 500 for complex dynamic workshops. Furthermore, the memory 520 may include both internal storage units and external storage devices of the global forward-looking intelligent scheduling device 500 for complex dynamic workshops. The memory 520 is used to store application software and various types of data installed on the global forward-looking intelligent scheduling device 500 for complex dynamic workshops, such as the program code of the global forward-looking intelligent scheduling device 500 for complex dynamic workshops. The memory 520 can also be used to temporarily store data that has been output or will be output. In one embodiment, the memory 520 stores a global forward-looking intelligent scheduler 540 for complex dynamic workshops. The global forward-looking intelligent scheduler 540 for complex dynamic workshops can be executed by the processor 510 to realize the global forward-looking intelligent scheduling method for complex dynamic workshops in the various embodiments of this application.

[0075] In some embodiments, processor 510 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in memory 520 or process data, such as executing a global forward-looking intelligent scheduling method for complex dynamic workshops.

[0076] In some embodiments, display 530 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. Display 530 is used to display information from the global forward-looking intelligent scheduling device 500 for complex dynamic workshops and to display a user interface for visualization. Components 510-530 of the global forward-looking intelligent scheduling device 500 for complex dynamic workshops communicate with each other via a system bus.

[0077] In one embodiment, when the processor 510 executes the global forward-looking intelligent scheduler 540 for complex dynamic workshops stored in the memory 520, the steps in the global forward-looking intelligent scheduling method for complex dynamic workshops described above are implemented.

[0078] This embodiment also provides a computer-readable storage medium storing a global forward-looking intelligent scheduler for complex dynamic workshops. When executed by a processor, this global forward-looking intelligent scheduler for complex dynamic workshops implements the following steps: The collected raw workshop data is processed according to industrial-grade data preprocessing rules to construct the real-time status feature vector of the current process. Based on the process route information of the process to be scheduled, extract the multi-step future process path information and encode it into a structured feature vector; The real-time state feature vector and the structured feature vector are weighted and fused to generate a fused state representation; The fused state representation is input into a preset intelligent reasoning model, and structured scheduling instructions are output. Structured scheduling instructions are sent to the workshop execution layer, and execution feedback information is continuously received. Global forward-looking dynamic scheduling is triggered based on the actual execution status and dynamic disturbance events.

[0079] This application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method. The computer-readable storage medium may include, but is not limited to, any type of disk, including floppy disks, optical disks, DVDs, CD-ROMs, microdrives, as well as magneto-optical disks, ROMs, RAMs, EPROMs, EEPROMs, DRAMs, VRAMs, flash memory devices, magnetic cards or optical cards, nanosystems (including molecular memory ICs), or any type of medium or device suitable for storing instructions and / or data.

[0080] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

[0081] In the above embodiments, the descriptions of each embodiment have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0082] In the embodiments provided in this application, it should be understood that the disclosed system can be implemented in other ways. For example, the system embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the shown or discussed mutual couplings, direct couplings, or communication connections may be through some service interfaces; indirect couplings or communication connections between systems or units may be electrical or other forms.

[0083] Those skilled in the art will understand that all or part of the steps in the various methods of the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, which may include: a flash drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, etc.

[0084] The foregoing description is merely an exemplary embodiment of this disclosure and should not be construed as limiting the scope of this disclosure. Any equivalent changes and modifications made in accordance with the teachings of this disclosure shall still fall within the scope of this disclosure. Those skilled in the art will readily conceive of embodiments of this disclosure upon considering the specification and practicing the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this disclosure that follow the general principles of this disclosure and include common knowledge or customary techniques in the art not described herein. The specification and embodiments are to be considered exemplary only, and the scope and spirit of this disclosure are defined by the claims.

[0085] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0086] Those skilled in the art will readily understand that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A global proactive intelligent scheduling method for complex dynamic shop floor, characterized in that, include: The collected raw workshop data is processed according to industrial-grade data preprocessing rules to construct the real-time status feature vector of the current process. Based on the process route information of the process to be scheduled, multi-step future process path information is extracted and encoded into a structured feature vector. Specifically, this includes: extracting all subsequent processes after the process to be scheduled based on the process route information to form a future process sequence; extracting the process identifier, optional processing equipment set, standard processing time, time offset relative to the current process, and demand intensity vector for various production resources from each subsequent process in the future process sequence as feature information for that subsequent process; constructing a two-dimensional future path information matrix with each subsequent process in the future process sequence as a row and the feature information corresponding to each subsequent process as a column; and mapping the two-dimensional future path information matrix into the structured feature vector using a pre-trained neural network encoder. The real-time state feature vector and the structured feature vector are weighted and fused to generate a fused state representation. Specifically, this includes: inputting the real-time state feature vector and the structured feature vector into a feature fusion network based on an attention mechanism; calculating dynamic attention weights for each feature dimension of the real-time state feature vector and the structured feature vector through the feature fusion network; and performing a weighted summation of the real-time state feature vector and the structured feature vector based on the dynamic attention weights, followed by a nonlinear transformation, to output the fused state representation. The fused state representation is input into a preset intelligent reasoning model, which outputs a structured scheduling instruction. The preset intelligent reasoning model is a large language model that has been fine-tuned by deep knowledge of the workshop scheduling domain. The input layer of the large language model receives the fused state representation, the weighted dependency matrix output by the dependency quantization unit, and the equipment load level characteristics output by the time series prediction unit. The weighted dependency matrix is ​​used to provide the large language model with structural constraint information arising from resource competition between processes; the equipment load level feature is used to provide the large language model with information on the load stress of key equipment in future time periods; The structured scheduling instructions are sent to the workshop execution layer, and execution feedback information is continuously received. Global forward-looking dynamic scheduling is triggered based on the actual execution status and dynamic disturbance events.

2. The globally proactive intelligent scheduling method for complex dynamic shop floor of claim 1, wherein, The process of preprocessing the collected raw workshop data according to industrial-grade data preprocessing rules also includes: The system determines whether the deviation between the original workshop data and the historical average exceeds a preset fluctuation threshold and smooths out noisy data. Abnormal data is marked and removed according to preset process rules, and historical statistical values ​​are used to fill in the missing data after removal; The filled data is transformed into a unified numerical range using a linear mapping method to generate a standardized snapshot of the workshop's global state.

3. The globally proactive intelligent scheduling method for complex dynamic shop floor of claim 2, wherein, The construction of the real-time state feature vector of the current process also includes: The status information of each selectable processing equipment for the workpiece corresponding to the current process, the number of waiting processing tasks in the equipment queue, and the estimated idle time of the equipment are obtained from the global status snapshot of the workshop. Obtain the current inventory level of materials required for the current process from the global snapshot of the workshop; The status information of the optional processing equipment, the number of waiting processing tasks, the estimated idle time, and the current inventory level are concatenated and vectorized to form the real-time status feature vector.

4. The globally proactive intelligent scheduling method for complex dynamic shop floor of claim 1, wherein, The step of issuing the structured scheduling instructions to the workshop execution layer and continuously receiving execution feedback information, and triggering global forward-looking dynamic scheduling based on the actual execution status and dynamic disturbance events, also includes: It continuously receives real-time production data from the workshop execution layer to monitor for any type of disturbance event, such as equipment failure, emergency order insertion, or material shortage, or whether the preset periodic scheduling and inspection time point has been reached. When a disturbance occurs or a periodic scheduling check point arrives, calculate the deviation between the earliest start time of each process and the baseline planned time, and estimate the impact of the disturbance on the total completion time or order delivery date. The degree of impact is compared with a preset impact threshold. If the degree of impact exceeds the preset impact threshold, a global forward-looking dynamic scheduling process is triggered to update the scheduling instruction. If the degree of impact does not exceed the preset impact threshold, the planned start time and planned end time of the affected process will be adjusted based on the original scheduling scheme.

5. A globally proactive intelligent scheduling system for complex dynamic shop floor, characterized in that, include: The preprocessing module is configured to process the collected raw workshop data according to industrial-grade data preprocessing rules to construct the real-time status feature vector of the current process. The encoding module is configured to extract multi-step future process path information and encode it into a structured feature vector based on the process route information of the process to be scheduled. Specifically, it includes: extracting all subsequent processes after the process to be scheduled based on the process route information to form a future process sequence; extracting the process identifier, optional processing equipment set, standard processing time, time offset relative to the current process, and demand intensity vector of various production resources from each subsequent process in the future process sequence as feature information of the subsequent process; constructing a two-dimensional future path information matrix with each subsequent process in the future process sequence as a row and the feature information corresponding to each subsequent process as a column; and mapping the two-dimensional future path information matrix into the structured feature vector through a pre-trained neural network encoder. The weighting module is configured to perform weighted fusion of the real-time state feature vector and the structured feature vector to generate a fused state representation. Specifically, it includes: inputting the real-time state feature vector and the structured feature vector into a feature fusion network based on an attention mechanism; calculating dynamic attention weights for each feature dimension of the real-time state feature vector and the structured feature vector through the feature fusion network; and performing a weighted summation of the real-time state feature vector and the structured feature vector based on the dynamic attention weights, followed by a nonlinear transformation, and outputting the fused state representation. The instruction generation module is configured to input the fused state representation into a preset intelligent reasoning model and output structured scheduling instructions. The preset intelligent reasoning model is a large language model that has been fine-tuned with in-depth knowledge of the workshop scheduling domain. The input layer of the large language model receives the fused state representation, the weighted dependency matrix output by the dependency quantization unit, and the equipment load level features output by the time series prediction unit. The weighted dependency matrix is ​​used to provide the large language model with structural constraint information arising from resource competition between processes; the equipment load level feature is used to provide the large language model with information on the load stress of key equipment in future time periods; The feedback adjustment module is configured to send the structured scheduling instructions to the workshop execution layer and continuously receive execution feedback information, triggering global forward-looking dynamic scheduling based on the actual execution status and dynamic disturbance events.

6. A globally proactive intelligent scheduling device for complex dynamic shop floor, characterized in that, It includes at least one processing unit and at least one storage unit, wherein the storage unit stores a computer program that, when executed by the processing unit, causes the processing unit to perform the steps of the global forward-looking intelligent scheduling method for complex dynamic workshops as described in any one of claims 1 to 4.

7. A storage medium, characterized by It stores a computer program that can be executed by a global forward-looking intelligent scheduling device for complex dynamic workshops. When the computer program is run on the global forward-looking intelligent scheduling device for complex dynamic workshops, it causes the global forward-looking intelligent scheduling device for complex dynamic workshops to perform the steps of the global forward-looking intelligent scheduling method for complex dynamic workshops as described in any one of claims 1 to 4.

Images