Cross-platform multi-modal industry chain intelligent scheduling method and device, and electronic equipment

Abstract

The application relates to the technical field of industrial chain intelligent scheduling, and discloses a cross-platform multi-modal industrial chain intelligent scheduling method and device and electronic equipment. The method collects cross-platform industrial chain data such as production, transportation, storage and demand, carries out multi-modal data fusion and semantic modeling, forms a unified semantic description network of the whole industrial chain, constructs a multi-layer digital twin model based on the fused data and the semantic network, generates a dynamic twin environment, triggers multi-agent distributed collaborative scheduling in the environment, generates a preliminary scheduling scheme, uses a discrete event simulation engine to carry out dynamic simulation, risk prediction and scheme optimization, and outputs a scheduling decision instruction set. The application can improve the collaborative utilization capability of cross-platform data, enhance the real-time performance, accuracy and anti-disturbance capability of industrial chain scheduling, and improve the operation efficiency, risk response capability and overall stability of the industrial chain in a complex scenario.

Description

Technical Field

[0001] This application relates to the field of intelligent scheduling technology for the industrial chain, specifically to a cross-platform multimodal intelligent scheduling method, device, and electronic equipment for the industrial chain. Background Technology

[0002] Supply chain scheduling systems face multiple structural obstacles in actual operation. Supply chain data sources are highly fragmented, covering equipment operation data in the production stage, logistics trajectory information in the transportation stage, inventory status records in the warehousing stage, and consumer behavior data on the market demand side. This data is independently managed by enterprises in different regions and stored in mutually isolated platform systems, forming data silos.

[0003] In related technologies, it is impossible to accurately identify the real interaction logic between cross-node entities, which leads to misjudgment of the overall operation status of the industrial chain. There is a lack of quantitative simulation mechanism for sudden shocks, making it difficult for the system to predict potential risk boundaries under uncertain operating conditions, resulting in a difficulty in balancing economic benefits and system stability. Summary of the Invention

[0004] This application provides a cross-platform, multimodal intelligent scheduling method, device, and electronic device for the industrial chain, which can solve the problems of data silos, slow response, and insufficient risk prediction in the prior art.

[0005] Firstly, this application provides a cross-platform multimodal intelligent scheduling method for the industrial chain. The method includes: performing data fusion processing on cross-platform industrial chain data to obtain multimodal fused data, and performing semantic modeling processing on the multimodal fused data to obtain a unified semantic description network for the entire industrial chain; constructing a multi-layer digital twin model for the entire industrial chain based on the multimodal fused data and the unified semantic description network for the entire industrial chain, and obtaining a dynamic twin environment based on the output of the multi-layer digital twin model for the entire industrial chain; triggering distributed collaborative scheduling operations of multiple agents based on the dynamic twin environment to generate a preliminary scheduling scheme; using the discrete event simulation engine inside the driving device to perform dynamic simulation and risk prediction on the preliminary scheduling scheme, generating a full-chain simulation evaluation report; extracting a comprehensive performance index sequence from the full-chain simulation evaluation report, and performing scheduling scheme optimization based on the comprehensive performance index sequence to obtain a set of scheduling decision instructions.

[0006] Based on the above technical means, the robustness of the system in uncertain environments is improved, thereby outputting scheduling decision instructions that balance efficiency and anti-interference.

[0007] Secondly, this application provides a cross-platform multimodal intelligent scheduling device for the industrial chain. The device includes: a data fusion unit, used to perform data fusion processing on cross-platform industrial chain data to obtain multimodal fused data, and to perform semantic modeling processing on the multimodal fused data to obtain a unified semantic description network for the entire industrial chain; a twin modeling unit, used to construct a multi-layer digital twin model for the entire industrial chain based on the multimodal fused data and the unified semantic description network for the entire industrial chain, and to obtain a dynamic twin environment based on the output of the multi-layer digital twin model for the entire industrial chain; an agent scheduling unit, used to trigger distributed collaborative scheduling operations of multiple agents based on the dynamic twin environment to generate a preliminary scheduling scheme; a simulation and deduction unit, used to use the preliminary scheduling scheme to drive the discrete event simulation engine inside the device to perform dynamic simulation and risk prediction on the preliminary scheduling scheme, and generate a full-chain deduction and evaluation report; and a decision optimization unit, used to extract a comprehensive performance index sequence from the full-chain deduction and evaluation report, and to perform scheduling scheme optimization based on the comprehensive performance index sequence to obtain a set of scheduling decision instructions.

[0008] Thirdly, this application provides an electronic device, including: a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the cross-platform multimodal supply chain intelligent scheduling method described in the first aspect or any corresponding embodiment.

[0009] This application provides a cross-platform, multimodal intelligent scheduling method, device, and electronic device for the industrial chain. It has the following beneficial effects: This application achieves a unified semantic description network for the entire industry chain by performing multimodal fusion processing on cross-platform industry chain data and further semantic modeling. This enables data from different platforms, such as production, transportation, warehousing, and demand, to be associated and expressed under a unified semantic framework. This reduces information fragmentation caused by inconsistencies in format, terminology, and semantics between different data sources, improves the accuracy of identifying entity relationships within the industry chain, and provides a more complete data foundation for subsequent digital twin modeling and scheduling decisions.

[0010] This application constructs a multi-layered digital twin model of the entire industry chain based on multimodal fusion data and a unified semantic description network, and generates a dynamic twin environment, enabling corresponding mappings of equipment, nodes, logistics paths, and operational status in the industry chain within the virtual environment. By combining real-time operational data, historical features, spatial topology, and entity behavior, the current state of the industry chain and its changing trends can be reflected more accurately, thereby improving the adaptability of scheduling schemes to actual operational scenarios.

[0011] This application triggers distributed collaborative scheduling of multiple agents in a dynamic twin environment and utilizes a discrete event simulation engine to dynamically simulate, predict risks, and optimize preliminary scheduling schemes. Each agent can form a preliminary scheduling scheme based on local state sharing and consistent updates. The simulation stage can evaluate indicators such as cost, latency, and inventory under disturbance conditions, thereby obtaining a set of scheduling decision instructions. This helps improve scheduling response speed and reduce the impact of sudden disturbances on the stability of the industrial chain. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0013] Figure 1 This is a schematic diagram of the first step of the cross-platform multimodal intelligent scheduling method for the industrial chain according to an embodiment of this application; Figure 2 This is a schematic diagram of the second process of the cross-platform multimodal intelligent scheduling method for the industrial chain according to an embodiment of this application; Figure 3 This is a schematic diagram of the third process of the cross-platform multimodal intelligent scheduling method for the industrial chain according to the embodiments of this application; Figure 4 This is a structural block diagram of a cross-platform multimodal intelligent scheduling device for the industrial chain according to an embodiment of this application; Figure 5 This is a comparison curve of the total cost of the entire chain under different scheduling strategies of this application as a function of a unified time step; Figure 6 This is a graph showing the change in global expected cumulative reward during the parameter iteration process of the deep neural network in this application.

[0014] Figure 7 This is a schematic diagram of the hardware structure of an electronic device according to an embodiment of this application.

[0015] Among them, 401 is the data fusion unit; 402 is the twin modeling unit; 403 is the agent scheduling unit; 404 is the simulation and deduction unit; 405 is the decision optimization unit; 701 is the processor; 702 is the ROM; 703 is the RAM; 704 is the bus; 705 is the I / O interface; 706 is the input device; 707 is the output device; 708 is the memory; and 709 is the communication device. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] It is understood that before using the technical solutions disclosed in the various embodiments of this application, users should be informed of the types, scope of use, and usage scenarios of the personal information involved in this application in an appropriate manner in accordance with relevant laws and regulations, and user authorization should be obtained.

[0018] The terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0019] In related technologies, supply chain scheduling systems face problems of data silos and insufficient semantic understanding when processing multi-source, heterogeneous, and cross-platform data, making it difficult to identify entity interaction relationships. Their centralized control architecture is prone to computational bottlenecks, resulting in long instruction transmission cycles, slow scheduling responses, and an inability to handle resource conflicts caused by unforeseen circumstances. Furthermore, existing solutions lack the ability to quantitatively extrapolate external random disturbances, making it difficult to predict risks and leading to insufficient robustness of the system in uncertain environments.

[0020] To address this, this application proposes a cross-platform, multimodal intelligent scheduling method for the industrial chain, such as... Figure 1 As shown, the method includes: Step S101: Perform data fusion processing on cross-platform industry chain data to obtain multimodal fusion data, and perform semantic modeling processing on multimodal fusion data to obtain a unified semantic description network for the entire industry chain.

[0021] Cross-platform supply chain data refers to heterogeneous data sets originating from different systems, platforms, or enterprise entities, covering multiple stages such as production, transportation, inventory, and market demand. The data may exist in various modalities such as text, numerical values, images, and sensor signals, reflecting the operational status of the supply chain.

[0022] Multimodal fusion data refers to the processing and integration of the aforementioned cross-platform industry chain data into a unified data representation with consistent format and semantics. Its purpose is to eliminate data silos and provide a consistent data foundation for subsequent semantic understanding and model building.

[0023] A unified semantic description network for the entire industry chain refers to a knowledge network constructed by semantically modeling multimodal fusion data, capable of uniformly describing entities, relationships, and attributes within the industry chain. Its aim is to provide a deep semantic understanding of the operational status of the industry chain.

[0024] Specifically, data from different sources and formats within the industry chain can be directly spliced ​​and format-converted manually or based on preset rules. For example, tabular data from different databases can be merged directly, or text data can be extracted for keywords and then correlated with numerical data. The processing method focuses on the physical integration of data, forming a raw, fused dataset containing multiple data types and sources.

[0025] Step S102: Based on multimodal fusion data and a unified semantic description network for the entire industry chain, construct a multi-layer digital twin model for the entire industry chain, and obtain a dynamic twin environment based on the output of the multi-layer digital twin model for the entire industry chain.

[0026] Among them, the multi-layered digital twin model of the entire industry chain refers to a virtual model built on fused data and semantic networks, capable of simulating the physical entities, operational processes, and behavioral logic of the industry chain. This model typically contains multiple layers to reflect the structure and function of the industry chain.

[0027] A dynamic twin environment refers to a virtual operating environment constructed based on the real-time output of a digital twin model and combined with actual operational data. This environment can reflect the current state and future evolution trends of the industry chain in real time. It supports dynamic perception and decision-making.

[0028] Specifically, semantic modeling is performed on multimodal fused data to obtain a unified semantic description network for the entire industry chain. Descriptions of entities in the industry chain, such as factories, warehouses, and transport vehicles, and their relationships, can be manually constructed based on expert experience or predefined ontology. For example, specific fields in the fused data can be mapped to preset entity and relationship types to form a preliminary graph structure. This relies on manual definition and maintenance, aiming to initially establish semantic connections within the industry chain.

[0029] Based on multimodal fusion data and a unified semantic description network across the entire industry chain, a multi-layered digital twin model of the entire industry chain is constructed. Using manual modeling software, two-dimensional schematic diagrams of each node in the industry chain can be drawn based on the physical layout and equipment list, and basic attribute information such as production capacity and storage capacity can be added to each node. This information can serve as the foundational layer of the digital twin model, reflecting the static structure of the industry chain.

[0030] Based on the output of the multi-layered digital twin model of the entire industry chain, a dynamic twin environment is obtained. Key operational data, such as production volume, inventory levels, and transportation status, can be periodically collected from the actual system and then directly updated to the corresponding node attributes in the aforementioned digital twin model. This allows the model to reflect the current state of the industry chain, with its dynamism primarily manifested in the periodic updates of the data.

[0031] Step S103: Based on the dynamic twin environment, trigger multi-agent distributed collaborative scheduling operations to generate a preliminary scheduling scheme.

[0032] In this context, multi-agent refers to multiple independent or semi-independent computing entities in a distributed system, each possessing certain perception, decision-making, and action capabilities. They collaborate and work together to accomplish complex tasks.

[0033] Distributed cooperative scheduling refers to the process by which multiple intelligent agents, without central control, jointly formulate and execute scheduling strategies through mutual communication and cooperation. It aims to improve the system's response speed and robustness.

[0034] The initial scheduling scheme refers to an initial, unoptimized scheduling plan generated by multiple agents through distributed cooperative scheduling operations in a dynamic twin environment. This scheme serves as input for subsequent simulations and optimizations.

[0035] Specifically, each key node in the industrial chain, such as a factory or warehouse, can be configured with an independent intelligent agent. Each agent independently formulates its own scheduling strategy based on the dynamic twin environment information it perceives locally. These agents exchange their current resource needs or availability information through point-to-point message passing, and then adjust their strategies to achieve resource allocation at the local level. Thus, each agent generates its own local scheduling scheme, and the collection of these local schemes constitutes the initial scheduling scheme.

[0036] Step S104: Using the discrete event simulation engine inside the preliminary scheduling scheme driving device, perform dynamic simulation and risk prediction on the preliminary scheduling scheme, and generate a full-chain simulation and evaluation report.

[0037] Discrete event simulation engine refers to a computational tool used to simulate system behavior. It infers the system's operating state and performance under different conditions by simulating the occurrence and processing of discrete events in the system.

[0038] A full-chain simulation and evaluation report is a detailed assessment document generated after using a discrete event simulation engine driven by a preliminary scheduling scheme to perform dynamic simulation and risk prediction. This report includes performance indicators, risk analysis, and predictions of potential problems for each link in the industry chain.

[0039] Specifically, a preliminary scheduling plan can be used as input and simulated in a discrete event simulation engine. During the simulation, some external disturbances, such as equipment failures or transportation delays, can be preset, and their impact on the execution of the scheduling plan can be observed. The simulation results can be recorded as a series of event logs to evaluate the plan's performance under specific disturbances. The event logs recorded during the simulation are then summarized and statistically analyzed, for example, to calculate total production time, number of transports, and inventory changes. These statistical data can be presented in text or tabular form as an evaluation result of the preliminary scheduling plan.

[0040] Step S105: Extract the comprehensive performance index sequence from the full-chain simulation and evaluation report, and optimize the scheduling scheme based on the comprehensive performance index sequence to obtain the scheduling decision instruction set.

[0041] The comprehensive performance index sequence refers to a series of quantitative indicators extracted from the full-chain simulation and evaluation report, used to measure aspects such as operational efficiency, cost, risk, and robustness of the industrial chain. These indicators are used to guide the optimization of scheduling schemes.

[0042] A scheduling decision instruction set refers to a set of executable instructions generated after optimization of the scheduling scheme based on a comprehensive performance index sequence. These instructions are used to guide the specific operations of various entities and equipment in the industry chain to achieve the optimization goals.

[0043] Specifically, a sequence of comprehensive performance indicators is extracted from the end-to-end performance evaluation report. Predefined key performance indicators, such as total cost, total time, and resource utilization, can be identified and extracted from the text or tabular data of the evaluation report, either manually or through automated scripts. These indicators are then ranked according to their importance, forming an indicator list.

[0044] Furthermore, based on the comprehensive performance index sequence, the scheduling scheme is optimized to obtain a set of scheduling decision instructions. The initial scheduling scheme can be adjusted based on the extracted comprehensive performance index sequence, either through manual analysis or based on preset optimization rules. For example, if a certain index is found to have failed to meet the preset performance target, a parameter in the scheme is modified, and then simulation evaluation is performed again until the preset performance target is achieved. The final adjustment result can be translated into a series of specific execution instructions.

[0045] It is understood that this application addresses the issues of data silos and insufficient semantic understanding by fusing and semantically modeling cross-platform industry chain data to construct a unified semantic description network, thereby achieving perception of the industry chain's operational status. By constructing a multi-layered digital twin model and a dynamic twin environment, and triggering distributed collaborative scheduling among multiple agents, the scheduling response speed is improved, avoiding the computational bottlenecks and resource conflicts inherent in centralized control. Furthermore, utilizing a discrete event simulation engine for dynamic simulation and risk prediction enables the quantitative assessment of potential risks, enhancing the system's robustness in uncertain environments, and ultimately outputting scheduling decision instructions that balance efficiency and anti-interference capabilities.

[0046] In some embodiments, this application further proposes a cross-platform multimodal intelligent scheduling method for the industrial chain, such as... Figure 2 As shown, the method includes: Step S201 involves performing data fusion processing on cross-platform industry chain data to obtain multimodal fused data, and then performing semantic modeling processing on the multimodal fused data to obtain a unified semantic description network for the entire industry chain. Specifically, this includes: Step S2011: Extract features from cross-platform industry chain data to generate independent feature vectors for each modality; Step S2012: Perform a linear transformation on the independent feature vectors based on the feature mapping weight matrix to generate feature representation matrices corresponding to each modality; Step S2013: Calculate the dynamic attention weight value of the feature representation matrix of each modality in the current system state based on the dynamic attention mechanism; Step S2014: Multiply the dynamic attention weight values ​​with the corresponding feature representation matrices, and perform a summation operation on the product results of all modalities to obtain multimodal fusion data; Step S2015: Perform semantic modeling processing on the multimodal fusion data to obtain a unified semantic description network for the entire industry chain.

[0047] Specifically, cross-platform supply chain data refers to heterogeneous data sets originating from different systems, organizations, and geographical locations, covering multiple stages such as production, transportation, warehousing, sales, and market demand. This data may exist in structured, semi-structured, or unstructured forms and is typically distributed across independent platforms or systems, forming data silos.

[0048] Feature extraction refers to identifying and extracting representative and discriminative information from raw data, transforming it into a more concise and easier-to-process numerical representation. Its purpose is to reduce data dimensionality, remove redundant information, and enhance the identifiability of data patterns, thereby laying the foundation for subsequent data fusion and analysis. For example, principal component analysis (PCA) and other statistical methods can be used to extract the main features of the data; or deep learning models can be used to process time series data, automatically learning and extracting high-level abstract features.

[0049] Independent feature vectors for each modality refer to numerical sequences that independently represent the core information of a modality, obtained after feature extraction from different types of original data. Each vector represents an abstract representation of specific modality data in a certain dimension. For example, for text data, it could be a word embedding vector; for sensor data, it could be a time-series feature vector after dimensionality reduction.

[0050] A feature mapping weight matrix is ​​a linear transformation matrix used to project feature vectors from different modalities onto a unified feature space. Its function is to adjust the scale, orientation, and importance of features from different modalities to eliminate heterogeneity and promote feature alignment. It can be learned and optimized through supervised or unsupervised learning.

[0051] Linear transformation refers to the operation of mapping points in one vector space to another through linear operations such as matrix multiplication. It is used to transform independent eigenvectors of various modes through a feature mapping weight matrix, making them comparable and fusionable in the new feature space.

[0052] The feature representation matrix is ​​the numerical matrix in a unified feature space corresponding to each modality after linear transformation. It contains the feature information of the modality data after standardization and alignment before fusion, providing a unified input format for subsequent dynamic attention mechanisms and summation operations.

[0053] Dynamic attention mechanisms are computational methods that adaptively allocate the importance of different modal features based on the current system state. Their purpose is to enable the system to focus on the most relevant and informative modal data, thereby improving the effectiveness and accuracy of data fusion. For example, a small neural network can take the current system state as input and output the attention weight for each modality; alternatively, a self-attention mechanism based on the Transformer architecture can be used to calculate the inter-modal dependencies through a query-key-value model and adjust it in conjunction with the system state.

[0054] The current system state refers to the real-time operational status and external environmental conditions of the supply chain system at a specific point in time. This includes, but is not limited to, the operating parameters of production equipment, inventory levels, order volume, transportation information in transit, market demand forecasts, weather conditions, and unforeseen events. This state information is an important basis for the dynamic attention mechanism to adjust modal weights.

[0055] The dynamic attention weights are the output of the dynamic attention mechanism, representing the contribution of the feature representation matrix of each modality to the final fusion result under the current system state. These weights are typically floating-point numbers between 0 and 1, reflecting the relative importance of different modal information in a specific context.

[0056] The summation operation on the product results involves multiplying the feature representation matrix of each modality by its corresponding dynamic attention weight value, and then summing the product results of all modalities. This achieves weighted fusion of information from different modalities, allowing modal information with higher attention weights to play a more dominant role in the fused data.

[0057] Multimodal fusion data contains comprehensive semantic information from all aspects of the industrial chain, and can more comprehensively and accurately reflect the overall operating status of the industrial chain.

[0058] Semantic modeling refers to the structuring and conceptualization of multimodal fused data to reveal the entities, relationships, and attributes inherent in the data, and to construct a knowledge representation model that can be understood and reasoned about by machines. Its role is to elevate the fused data from a purely numerical representation to a level with explicit semantics, thereby supporting higher-level analysis and decision-making. For example, knowledge graphs can be constructed to represent entities in the industry chain, such as factories, warehouses, and products, and their interrelationships, such as production, transportation, and storage, in graph form.

[0059] A unified semantic description network for the entire industry chain can clearly depict the logical relationships, business processes, and resource flows between various nodes in the industry chain (such as manufacturers, suppliers, and logistics companies), providing a comprehensive, accurate, and reasonable semantic foundation for intelligent scheduling.

[0060] In one example, the cross-platform industry chain data is cleaned to remove invalid and null values. After cleaning, the cross-platform industry chain data is classified and standardized according to preset metadata management rules. For the standardized multi-source time-series data, the corresponding temporal and spatial features are extracted, and spatiotemporal alignment and normalization calculations are performed to generate the processed standard data.

[0061] The specific formulas used to perform spatiotemporal alignment and normalization calculations for multi-source time series data are as follows: ; in, For a specific moment; Index the data source; This is the time window offset; The raw data includes time window offsets; This is a time interpolation alignment function; and Data sources Historical mean and standard deviation; This is the processed standard data.

[0062] In one example, after extracting features from cross-platform industry chain data and generating independent feature vectors for each modality, the process also includes performing multimodal feature fusion calculation based on a dynamic attention mechanism, as shown in the following formula: ; in, This represents the total number of data modalities. For modal indexing; For modality Independent eigenvectors; For modality The feature mapping weight matrix; For modality Dynamic attention weights; This is the output cross-modal fusion feature vector.

[0063] Step S202: Based on multimodal fusion data and a unified semantic description network across the entire industry chain, a multi-layered digital twin model of the entire industry chain is constructed, and a dynamic twin environment is obtained based on the output of the multi-layered digital twin model of the entire industry chain. (See details...) Figure 1 Step S102 in the embodiment will not be described again here.

[0064] Step S203: Based on the dynamic twin environment, trigger multi-agent distributed cooperative scheduling operations to generate a preliminary scheduling scheme. See details... Figure 1 Step S103 in the embodiment will not be described again here.

[0065] Step S204: Utilize the discrete event simulation engine within the preliminary scheduling scheme driving device to perform dynamic simulation and risk prediction on the preliminary scheduling scheme, generating a full-chain simulation and evaluation report. See details... Figure 1 Step S104 in the embodiment will not be described again here.

[0066] Step S205: Extract the comprehensive performance index sequence from the full-chain simulation and evaluation report, and based on the comprehensive performance index sequence, perform scheduling scheme optimization to obtain a set of scheduling decision instructions. See details in the following section. Figure 1 Step S105 in the embodiment will not be described again here.

[0067] It is understood that, through the above-described technical solution in this embodiment, this application effectively solves the problems of existing methods in processing cross-platform industry chain data, such as merely stacking surface numerical values, lacking deep semantic parsing mechanisms, resulting in large feature alignment errors, and difficulty in accurately identifying the real interaction relationships between cross-node entities. Specifically, firstly, by extracting features from the cross-platform industry chain data, independent feature vectors for each modality are generated. This provides a denoising and refined basic representation for subsequent fusion processing, avoiding the complexity and redundancy of the original data. Secondly, a linear transformation is performed on the independent feature vectors based on the feature mapping weight matrix to generate feature representation matrices corresponding to each modality. This projects the features of different modalities onto a unified feature space, effectively reducing the heterogeneity between modalities and lowering feature alignment errors. Furthermore, based on a dynamic attention mechanism, the dynamic attention weight value of the feature representation matrix of each modality in the current system state is calculated, and then multiplied with the corresponding feature representation matrix and summed to obtain multimodal fusion data. This dynamic weighted fusion process enables the system to adaptively adjust the contribution of different modal information according to the real-time operation of the industrial chain, ensuring that key information receives priority attention during the fusion process. This overcomes the limitations of static fusion methods and achieves deep semantic analysis of multimodal heterogeneous data. Finally, semantic modeling processing is performed on the multimodal fusion data to obtain a unified semantic description network for the entire industrial chain. This network clearly depicts entities, relationships, and attributes in the industrial chain in a structured form, enabling the system to accurately identify the real interaction relationships between cross-node entities and comprehensively and accurately understand the overall operation state of the industrial chain. Through the above series of steps, this application realizes the transformation from data silos to a unified semantic description network, providing high-quality, high-semantic-density input for the subsequent construction of a multi-layered digital twin model of the entire industrial chain, and improving the accuracy and robustness of intelligent scheduling of the industrial chain.

[0068] In some embodiments, this application further proposes performing semantic modeling processing on multimodal fusion data to obtain a unified semantic description network for the entire industry chain, specifically including: Step a1 involves performing a unified semantic mapping on the multimodal fusion data based on the entity relationship attribute model. The entity identifiers in the multimodal fusion data are used to match the corresponding industry chain entity nodes.

[0069] Entity-relationship-attribute model is a structured data model used to describe real-world entities, the relationships between entities, and the attributes that entities possess. It provides a unified semantic framework for heterogeneous data, enabling data from different modalities and sources to be consistently understood and processed.

[0070] Entity identifiers are tags or codes used to uniquely identify entity nodes in a supply chain, such as equipment numbers, factory IDs, and product SKUs.

[0071] Specifically, semantic alignment and integration of data can be achieved by matching entity identifiers in multimodal fusion data with predefined industry chain entity nodes. In practice, semantic web technology can be used to construct entity relationship attribute models. For example, entity types such as "factory," "equipment," "product," and "transportation route" can be defined, and attributes such as factory location, equipment model, and product batch can be defined for them, along with relationships such as "factory produces products" and "equipment is located in a factory." After data fusion, the system parses the entity identifiers in the fused data and matches them with entities defined in the ontology, thereby mapping the original data to a unified semantic structure. Alternatively, the schema definition of relational databases or graph databases can be used to construct entity relationship attribute models. For example, in a graph database, various elements in the industry chain, such as suppliers, manufacturers, logistics companies, warehouses, and products, can be defined as nodes, and the connections between them can be defined as edges, with attributes attached to nodes and edges. The entity identifiers in the multimodal fusion data then serve as query keys to find and associate corresponding entity nodes in the graph database, achieving a unified semantic mapping.

[0072] Step a2: Obtain the list of attributes contained in the entity nodes of the industrial chain, and extract the knowledge graph embedding vector of the entity nodes of the industrial chain.

[0073] The attribute list is a series of key-value pairs that describe the characteristics of entity nodes in the industry chain. For example, a factory entity may contain attributes such as "geographical location", "capacity", and "main products".

[0074] Knowledge graph embedding vectors are representations that map entities and relations in a knowledge graph to a low-dimensional continuous vector space. They can capture the semantic information and structural features of entities and relations, enabling similarity calculation and reasoning in the vector space.

[0075] Attribute lists can be obtained by querying predefined database schemas, configuration files, or extracting them from structured or semi-structured data sources using a data parser. Knowledge graph embedding vectors can be extracted using knowledge graph embedding algorithms. Furthermore, attribute lists can also be extracted from unstructured text data using Natural Language Processing (NLP) techniques, such as identifying key information as attributes from equipment manuals and production reports. Knowledge graph embedding vector extraction can also leverage neural network-based methods to capture local and global information within the graph structure, generating more expressive entity embeddings.

[0076] Step a3: Obtain the relationship projection matrix associated with the entity nodes in the industrial chain based on the preset relationship type list.

[0077] The relation type list defines various semantic relationships that may exist between entities in the industry chain, such as "production", "supply", "transportation", "storage", and "dependency".

[0078] A relation projection matrix is ​​a transformation matrix used to project entity embedding vectors from their original space to a specific relation space. Each relation type corresponds to a unique projection matrix that captures the semantic characteristics of that relation type, enabling a more accurate assessment of the entity's relevance under that relation during relation reasoning.

[0079] The pre-defined list of relationship types can be defined by domain experts based on the actual business logic and data characteristics of the industry chain. Obtaining the relationship projection matrix is ​​typically part of the knowledge graph embedding algorithm training process. Alternatively, the relationship type list can be automatically discovered through cluster analysis or topic modeling of large amounts of industry chain data. The relationship projection matrix can also be obtained through deep learning models, such as those based on the Transformer architecture. These models can learn more complex contextual dependencies, thereby generating more refined relationship representations and projection matrices.

[0080] Step a4: Multiply the knowledge graph embedding vector with the relation projection matrix to generate entity projection vectors mapped to the corresponding relation space.

[0081] Specifically, this step combines the semantic representation of the entity—the knowledge graph embedding vector—with the semantic representation of the specific relation type—the relation projection matrix—to obtain a vector representation of the entity's "perspective" or "state" within that specific relation. This entity projection vector is no longer a representation of the entity in a general semantic space, but rather its semantic representation within the context of a specific relation, which is crucial for accurately calculating the strength of relations between entities. The multiplication operation can be standard matrix multiplication. Besides direct matrix multiplication, more complex nonlinear transformations can be employed, such as implementing projection through a small neural network layer, where the projection matrix is ​​the weight of that neural network layer. This can capture more complex projection relationships and improve the expressive power of the projection.

[0082] Step a5: Calculate the semantic relevance score of the entities in the industry chain based on the generated entity projection vector, and connect the entities with semantic relevance scores greater than the preset relevance score threshold to form an edge, thereby obtaining a unified semantic description network for the entire industry chain.

[0083] Specifically, the semantic relevance score is a quantitative indicator that measures the degree of semantic relevance between two entities in a specific industry chain under a particular relationship. This score is derived by comparing the similarity between the entity projection vectors. A preset relevance score threshold is a critical value used to determine whether a valid connection exists between entities. Only when the semantic relevance score between two entities exceeds this threshold is a sufficiently strong semantic connection considered to exist, thus establishing an edge in the unified semantic description network for the entire industry chain. The semantic relevance score can be calculated using various similarity measurement methods, such as cosine similarity, negative Euclidean distance, and dot product. The preset relevance score threshold can be determined based on historical data analysis, expert experience, or through methods such as cross-validation. The relevance score can also be calculated using methods based on energy functions or scoring functions. The threshold setting can also be dynamic, for example, adjusted according to the real-time operating status or scheduling needs of the industry chain. Finally, entities that meet the conditions are connected into edges to construct a graph structure that accurately reflects the semantic relationships between entities in the industry chain—that is, the unified semantic description network for the entire industry chain.

[0084] In one example, the specific formula for calculating the semantic relevance score of entities in the industry chain is as follows: ; in, and They represent two different entities in the industrial chain; and Entities With entity Knowledge graph embedding vectors; For a specific relation projection matrix; This is the function for calculating cosine similarity. Score the semantic relevance between entities.

[0085] It is understood that, through the above-described technical solutions in this embodiment, this application provides a refined semantic modeling processing mechanism, effectively addressing the shortcomings of multimodal fusion data in semantic unified mapping, knowledge representation extraction, and semantic relevance calculation. Specifically, semantic unified mapping is performed based on an entity relationship attribute model, and precise matching is achieved using entity identifiers, ensuring semantic consistency of data from different sources and modalities, fundamentally avoiding ambiguity and feature alignment errors caused by data heterogeneity. By obtaining the attribute list of entity nodes in the industry chain and extracting their knowledge graph embedding vectors, this application can deeply capture the inherent semantic and structural features of entities, providing rich knowledge representations for subsequent relationship reasoning. Furthermore, by obtaining a relationship projection matrix through a preset relationship type list and projecting the knowledge graph embedding vectors onto a specific relationship space, precise quantification of the relationship strength between entities is achieved. Finally, based on the calculated semantic relevance score and combined with preset threshold filtering, a highly accurate and semantically rich unified semantic description network for the entire industry chain is constructed. This provides a semantic foundation for the subsequent construction of a multi-layered digital twin model across the entire industry chain, improves the accuracy of the digital twin model and the effectiveness of simulation and inference, and thus ensures the decision-making quality of multi-agent distributed collaborative scheduling operations and the robustness and reliability of the overall scheduling method.

[0086] In some optional embodiments, this application further proposes a cross-platform multimodal intelligent scheduling method for the industrial chain, such as... Figure 3 As shown, the method includes: Step S301 involves performing data fusion processing on cross-platform industry chain data to obtain multimodal fused data, and then performing semantic modeling processing on the multimodal fused data to obtain a unified semantic description network for the entire industry chain. Specifically, this includes: Step S3011 involves extracting features from the cross-platform industry chain data to generate independent feature vectors for each modality. See details in the following section. Figure 2 Step S2011 in the embodiment will not be described again here.

[0087] Step S3012 involves performing a linear transformation on the independent feature vectors based on the feature mapping weight matrix to generate feature representation matrices corresponding to each modality. See details... Figure 2 Step S2012 in the embodiment will not be described again here.

[0088] Step S3013: Calculate the dynamic attention weight values ​​of the feature representation matrix of each modality in the current system state based on the dynamic attention mechanism. See details... Figure 2 Step S2013 in the embodiment will not be described again here.

[0089] Step S3014 involves multiplying the dynamic attention weights by the corresponding feature representation matrices and summing the product results across all modalities to obtain the multimodal fusion data. (See details...) Figure 2 Step S2014 in the embodiment will not be described again here.

[0090] Step S3015 involves performing semantic modeling processing on the multimodal fusion data to obtain a unified semantic description network for the entire industry chain. See details in the following section. Figure 2 Step S2015 in the embodiment will not be described again here.

[0091] Step S302: Based on multimodal fusion data and a unified semantic description network across the entire industry chain, construct a multi-layered digital twin model for the entire industry chain. Specifically, this includes: Step S3021: Extract the structural dimension parameters from the multimodal fusion data and generate the corresponding three-dimensional mesh model based on the structural dimension parameters.

[0092] Specifically, the process begins by identifying and extracting quantitative information describing the geometry, dimensions, spatial coordinates, and connection points of various physical entities within the industrial chain, such as production equipment, warehousing facilities, transportation vehicles, and factory buildings, from the fused multimodal data. This information constitutes structural dimension parameters. For example, these parameters can be obtained by parsing CAD drawings, BIM models, laser scanning point cloud data, high-precision photogrammetry data, or sensor array data. Secondly, based on these extracted structural dimension parameters, computer graphics and geometric modeling techniques are used to construct a three-dimensional digital model—a three-dimensional mesh model—that accurately represents the appearance and internal structure of each physical entity. This can be achieved through semi-automated modeling using parametric modeling software, by converting discrete point cloud data into a continuous mesh model using point cloud data reconstruction algorithms, or by matching and calling existing 3D model libraries. The aim is to provide an accurate physical structural foundation for digital twin models, avoiding model distortion caused by inaccurate structural information.

[0093] Step S3022: Assemble and splice the three-dimensional mesh model according to its spatial relative position and connection relationship to form a spatial topology framework for the entire industry chain.

[0094] Specifically, after generating individual 3D mesh models, they need to be integrated into a unified 3D virtual space based on their actual spatial layout in the real-world industrial chain and their physical or logical connections. "Spatial relative position" refers to the distance, orientation, and posture information between the entity models; "connection relationship" refers to the physical connections between entities, such as pipes, conveyor belts, assembly interfaces, or functional associations, such as upstream and downstream processes and logistics paths. The assembly and splicing process can be achieved through defining coordinate transformation matrices between models, utilizing assembly constraints such as coplanarity, coaxiality, and distance constraints for automated alignment and combination, or through modular combination based on a predefined topological relationship map. The final spatial topological framework of the entire industrial chain is a 3D digital skeleton that accurately reflects the overall layout of the industrial chain and the spatial relationships between its components, providing an accurate carrier for subsequent physical flow and information flow simulations.

[0095] Step S3023: Configure the corresponding rated operating parameter matrix for each node in the spatial topology framework of the entire industry chain, establish the correspondence between the actual equipment and the spatial topology framework of the entire industry chain, obtain the basic mapping model between the actual equipment and the system attributes, and use the basic mapping model as a multi-layer digital twin model.

[0096] Specifically, after constructing the spatial topology framework, each virtual node in the framework needs to be assigned performance attributes relevant to its actual operation. The rated operating parameter matrix is ​​a set of key performance indicators and attributes describing these nodes under normal operating conditions. For example, for production equipment, this might include rated capacity, energy consumption, processing accuracy, failure rate, and maintenance cycle; for warehousing nodes, it might include maximum storage capacity, throughput, and temperature and humidity control range. These parameters can be configured through database association, configuration file import, or API interface acquisition. Simultaneously, a one-to-one correspondence needs to be established between physical devices in the real world and corresponding virtual nodes in the digital twin model. This is typically achieved through unique identifiers, such as device IDs or asset codes, or attribute matching based on semantic networks. By integrating structural information, spatial topology information, and operating parameter information, a basic mapping model between actual equipment and system attributes is formed. This model can be stored and managed using relational databases, object-oriented models, knowledge graphs, or data models from industrial IoT platforms, serving as the foundational layer of the multi-layered digital twin model. This allows the digital twin model to not only have a physical appearance but also behavioral attributes, enabling it to simulate the operating state of equipment and the dynamic response of the system.

[0097] Step S303: Based on the dynamic twin environment, trigger multi-agent distributed cooperative scheduling operations to generate a preliminary scheduling scheme. See details... Figure 1Step S103 in the embodiment will not be described again here.

[0098] Step S304: Utilize the discrete event simulation engine within the preliminary scheduling scheme driving device to perform dynamic simulation and risk prediction on the preliminary scheduling scheme, generating a full-chain simulation and evaluation report. See details... Figure 1 Step S104 in the embodiment will not be described again here.

[0099] Step S305: Extract the comprehensive performance index sequence from the full-chain simulation and evaluation report, and based on the comprehensive performance index sequence, perform scheduling scheme optimization to obtain a set of scheduling decision instructions. See details in the following section. Figure 1 Step S105 in the embodiment will not be described again here.

[0100] It is understood that, through the above-described technical solution in this embodiment, this application further constructs an accurate multi-layered digital twin model of the entire industry chain based on multimodal fusion data and a unified semantic description network for the entire industry chain, solving the problem that digital twin models cannot truly reflect the actual structure and operating status of the industry chain. Specifically, by extracting structural dimension parameters from the multimodal fusion data and generating corresponding three-dimensional mesh models, it ensures that the basic construction of the model is based on the structural information in the actual data, avoiding feature alignment errors caused by data silos or insufficient semantic understanding, thereby accurately capturing the physical structure of the industry chain. Subsequently, the generated three-dimensional mesh models are assembled and spliced ​​according to spatial relative positions and connection relationships to form a spatial topology framework for the entire industry chain. This ensures that the topology of the model accurately reflects the real spatial layout and connection relationships of the industry chain, preventing assembly errors or omissions. Finally, corresponding rated operating parameter matrices are configured for each node in the spatial topology framework, and a correspondence between actual equipment and the framework is established, resulting in a basic mapping model of actual equipment and system attributes as a digital twin model. This allows the model to not only contain structural information but also incorporate operating parameters and equipment attributes, thereby comprehensively simulating the operating status of the industry chain and solving the model disconnect problem caused by missing equipment parameters. This provides a foundation for the subsequent generation of a dynamic twin environment, enabling scheduling operations based on this environment to be more accurate and efficient, thereby improving the reliability and decision-making quality of intelligent scheduling across the entire industry chain.

[0101] In some optional embodiments, this application further proposes obtaining a dynamic twin environment based on the output of a multi-layered digital twin model across the entire industry chain, specifically including: Step b1 involves binding the multimodal fusion data to the corresponding basic mapping model to establish a connection between the real-time running data layer and historical features. Step b2: Establish the causal feedback loop and logistics conservation equation between different industrial chain entity nodes, and define the system of differential equations for the state evolution of the industrial chain system. Step b3 involves overlaying and numerically fusing the 3D mesh model, entity behavior model, geographic information system data, and multi-agent parameters to obtain a dynamic twin environment.

[0102] Specifically, multimodal fusion data is bound to a corresponding basic mapping model to connect the real-time operational data layer with historical features. This aims to associate real-time collected and fused supply chain data with a pre-built basic mapping model representing the physical structure and attributes of the supply chain. For example, data stream processing technology can be used to match and update the real-time inflow of multimodal fusion data with the corresponding entity attributes in the basic mapping model through a preset data interface or API. Alternatively, an event-driven mechanism can be adopted, triggering a data synchronization service when the multimodal fusion data changes, updating the basic mapping model with the latest data. This ensures that the basic mapping model can reflect the current operational status of the supply chain in real time and is analyzed in conjunction with historical data features.

[0103] Establishing causal feedback loops and logistics conservation equations among different entity nodes in the industrial chain, and defining a set of differential equations for the state evolution of the industrial chain system, is to accurately characterize the dynamic interactions and material flow patterns among the entities within the industrial chain. Causal feedback loops can be represented by a directed graph model, where nodes represent entities in the industrial chain, such as production equipment, warehouses, and transportation routes, and edges represent information flow, material flow, or control flow between them, defining the direction and intensity of these flows. The logistics conservation equations, based on physical conservation laws, define the input, output, and internal state of each entity node, such as the dynamic balance relationship between inventory levels; for example, the inventory change rate of a warehouse equals its inbound rate minus its outbound rate. Furthermore, by transforming these causal feedback relationships and conservation equations into a series of interrelated differential equations, a complete set of differential equations for the state evolution of the industrial chain system can be formed, used to describe the continuous dynamic behavior of the entire industrial chain system over time.

[0104] A dynamic twin environment is obtained by overlaying and numerically fusing 3D mesh models, entity behavior models, geographic information system (GIS) data, and multi-agent parameters. This involves integrating multiple heterogeneous data sources into a unified virtual environment. Layer overlay can be achieved through a visualization engine, visualizing the 3D mesh model representing the physical spatial structure, entity behavior models describing the operational logic of equipment or personnel, GIS data providing geospatial background information such as road networks and regional boundaries, and multi-agent parameters defining the decision-making logic and interaction rules of agents as different information layers in the virtual space. Numerical fusion involves standardizing and normalizing these heterogeneous data at the data level, and integrating them into a unified mathematical model or simulation engine through specific fusion algorithms, such as weighted averaging, interpolation, and machine learning fusion. This forms a high-fidelity, computable, and interactive dynamic twin environment, providing a comprehensive real-time simulation foundation for subsequent scheduling decisions.

[0105] In one example, the specific formulas used in the system of differential equations are as follows: ; in, for The system's global state matrix at time 1; This is the system's endogenous state transition matrix; for The global scheduling instruction matrix at any given time; For scheduling control gain matrix; This is a matrix of external environment variables; This is the coupling matrix for environmental influences.

[0106] In one example, a simulation process with discrete time steps is initiated in the generated dynamic twin environment. The simulation process continuously records the state output values ​​of each entity node at the current unified time step.

[0107] Extract the actual output state values ​​of the real system collected by the sensing terminal within the same time window. Calculate the distance error between the simulated output values ​​of the dynamic twin environment and the actual output state values ​​of the real system. Calculate the weighted summation of the calculated distance error using an error diagonal penalty matrix to generate and output the system fidelity index of the digital twin spatiotemporal mapping. The specific formula used to evaluate the fidelity of the digital twin spatiotemporal mapping is as follows: ; in, This represents the total number of sampling observation steps. Index for observation steps; For real systems in Step output; For twin models in Simulation output of the step; This is the error diagonal penalty matrix; This is the system fidelity index.

[0108] It is understood that, through the above-described technical solution in this embodiment, this application binds multimodal fusion data with the basic mapping model, realizing the connection between the real-time operational data layer and historical features. This ensures that the digital twin model can acquire and integrate the latest operational data of the industrial chain in real time, avoiding the model update lag problem caused by data silos, thereby improving the timeliness and accuracy of the twin environment. Simultaneously, by establishing causal feedback loops, logistics conservation equations, and defining a set of differential equations for the state evolution of the industrial chain system, this application can accurately capture the dynamic interactions and material flow patterns among entities within the industrial chain. This allows the twin environment to not only reflect the static structure but also simulate the continuous dynamic evolution process of the industrial chain, enhancing its adaptability and predictive capabilities against sudden disturbances. Furthermore, by overlaying and numerically fusing 3D mesh models, entity behavior models, geographic information system data, and multi-agent parameters, a deep integration of multi-source heterogeneous information was achieved, constructing a comprehensive and high-fidelity dynamic twin environment. This provides a foundation for subsequent distributed collaborative scheduling operations triggered by multiple agents based on this environment, improving the response speed and risk prediction capabilities of the scheduling scheme, and effectively solving the problems of lack of dynamism, delayed updates, and difficulty in accurately reflecting real-time changes in the industrial chain.

[0109] In some embodiments, this application further proposes, based on a dynamic twin environment, triggering multi-agent distributed cooperative scheduling operations to generate a preliminary scheduling scheme, including: Step c1: Construct a multi-agent distributed communication network based on the spatial topology framework in the dynamic twin environment, and establish a local state sharing information channel based on the distributed communication network; Step c2: Perform the multi-agent cooperative network state consistency update operation. Based on the difference between the current agent's resource belief state and the resource belief state of the neighboring agents, combined with the communication topology adjacency weight and the consistency convergence step size, calculate and update the resource belief state for the next cooperative iteration number. Step c3: Activate the local decision optimization procedure, calculate the local net utility value corresponding to each local scheduling strategy based on the current updated resource belief state, and select the local scheduling strategy with the largest local net utility value as the initial scheduling scheme for the current agent.

[0110] Specifically, a multi-agent distributed communication network is constructed based on a spatial topology framework within a dynamic twin environment. The spatial topology framework is an abstract representation of the spatial structure and connections in the physical world, such as entities like production nodes, warehousing nodes, and transportation routes in a supply chain, and their interconnections. The purpose of constructing this multi-agent distributed communication network is to establish a decentralized information exchange mechanism for multiple agents, such as software agents representing production lines, transportation fleets, and inventory management systems. This provides a basic communication link for information sharing and collaborative decision-making among agents, avoiding the bottlenecks of traditional centralized communication. This communication network can achieve point-to-point or multicast message passing between agents by deploying virtual communication agents on the nodes and edges of the spatial topology framework and utilizing technologies such as message queues or distributed hash tables.

[0111] Building upon this foundation, a local state sharing information channel is established based on a distributed communication network. This channel provides a mechanism for each agent within the distributed communication network, enabling them to efficiently broadcast or query their own local state information (such as resource usage, task progress, and available capabilities) to or receive state information from neighboring agents. Its purpose is to ensure that agents can access directly relevant local environmental information during collaborative decision-making, providing a data foundation for subsequent state consistency updates. This channel can be implemented using a publish / subscribe model, where each agent subscribes to the state update topics of its neighboring agents and publishes its own state updates to the corresponding channel. Alternatively, a shared memory or distributed caching approach can be used, where agents write their local state to a shared area and periodically read the states of their neighboring agents.

[0112] Furthermore, a state consistency update operation is performed in the multi-agent cooperative network. This operation refers to the process in a multi-agent system where, through iterative computation, all or some agents reach a consensus on the state of shared resources they are interested in. Its purpose is to resolve resource conflicts that may arise from information asymmetry and local decision-making in distributed systems, ensuring that agents' perceptions of resource states tend to be consistent, thus providing a unified decision-making basis for collaborative scheduling. An average consensus algorithm can be used, where agents weight their resource belief states and the states of their neighbors, gradually converging to a consistent state. Alternatively, an iterative process based on distributed optimization algorithms can be employed, where each agent, while locally optimizing its resource allocation, exchanges information with neighboring agents, gradually adjusting its belief state until global consistency is achieved.

[0113] Based on the difference between the current resource belief state of the agent and that of its neighbors, and considering the adjacency weights of the communication topology and the consensus convergence step size, the resource belief state for the next cooperative iteration is calculated and updated. The resource belief state represents the agent's perception of the resources it manages or is interested in. By comparing its own belief state with that of its neighbors, and considering the communication topology and convergence step size, the agent can gradually adjust its perception of resource states, converging it towards the perceptions of its neighbors, ultimately achieving consensus. This enables information synchronization and consensus in the distributed system, providing a unified and accurate resource view for subsequent local decisions.

[0114] Subsequently, the local decision optimization procedure is activated. This procedure is an algorithmic module where each agent independently evaluates and selects the optimal course of action based on its current local information and updated resource belief state. This enables each agent to autonomously make the optimal scheduling decision for itself on a distributed and collaborative basis, thus forming a preliminary local scheduling strategy. This procedure can employ heuristic algorithms, such as greedy algorithms, genetic algorithms, or simulated annealing algorithms. The agents search for the optimal local scheduling strategy within a limited search space according to preset rules and an objective function.

[0115] After activating the local decision optimization procedure, the local net utility value corresponding to each local scheduling strategy is calculated based on the updated resource belief state. The local net utility value is a quantitative indicator that measures the effectiveness of a local scheduling strategy, typically considering both the benefits and costs of that strategy. Its purpose is to provide an objective evaluation standard for the agent to select the optimal local scheduling strategy.

[0116] Finally, the local scheduling strategy with the largest local net utility value is selected as the initial scheduling scheme for the current agent. This is the final output of the local decision optimization procedure. After evaluating the net utility values ​​of all possible local scheduling strategies, each agent selects the strategy with the highest net utility value as its own initial scheduling scheme. This ensures that each agent, given the current information and consensus state, can make the most advantageous decision for itself; the set of locally optimal decisions constitutes the initial scheduling scheme for the entire system. Agents can maintain a list of candidate local scheduling strategies, calculate the local net utility value for each strategy in the list, and then select the strategy with the largest net utility value through a simple comparison operation. Alternatively, a sorted or priority queue-based data structure can be used to store and sort the calculated local scheduling strategies and their net utility values, and then directly retrieve the strategy at the head of the queue.

[0117] In one example, the specific formula used for network state consistency updates is as follows: ; in, Index for the current agent; Indexing neighboring intelligent agents; For intelligent agents The set of communicating neighbors; This represents the number of collaborative iterations. For intelligent agents In the The next state of resource belief; For Neighbor Intelligent Agent In the The next state of resource belief; For communication topology adjacency weights; The step size for consistent convergence; For intelligent agents In the This is a state of resource belief.

[0118] It is understood that, through the above technical solution, this application, based on a dynamic twin environment, achieves decentralized and efficient information exchange between agents by constructing a multi-agent distributed communication network and a local state sharing information channel, effectively avoiding the bottleneck of centralized communication and thus accelerating information transmission and scheduling response. Simultaneously, a multi-agent cooperative network state consistency update operation is performed. Agents iteratively update their resource belief states based on differences in resource belief states with their neighbors, communication topology adjacency weights, and consistency convergence step size, ensuring that all agents reach a consensus on the shared resource state and effectively preventing resource overload and path conflicts caused by information asymmetry. A local decision optimization procedure is activated. Each agent, based on the updated consistent resource belief state, calculates and selects the scheduling strategy with the highest local net utility value, enabling each agent to make optimal local decisions under a unified and accurate resource view, thereby generating a preliminary scheduling scheme that minimizes conflict and optimizes efficiency.

[0119] In some optional embodiments, this application further proposes to utilize the discrete event simulation engine within the preliminary scheduling scheme driving device to perform dynamic simulation and risk prediction on the preliminary scheduling scheme, generating a full-chain deduction and evaluation report, specifically including: Step d1: Set up a dynamic disturbance variable generation program to simulate random changes in external environmental inputs and construct a composite injection model of random disturbance intensity and occurrence probability.

[0120] The dynamic disturbance variable generation program is a software module or algorithm used to simulate uncertainties in the external environment. Its function is to introduce randomness, making the simulation environment closer to various unforeseen situations that may occur in the real world. For example, this program can generate random variables such as market demand fluctuations, equipment failure rates, and weather changes based on historical data or preset probability distributions. The composite injection model of random disturbance intensity and probability aims to quantify and combine the impact of different types of disturbances and their likelihood of occurrence, considering the relationship between the magnitude of the disturbance and its frequency of occurrence. For example, a Bayesian network model can be constructed, using the intensity and probability of different disturbance events as nodes, and describing their mutual influence through conditional probability.

[0121] In one example, the specific formula used to construct the composite injection model of random perturbation strength and occurrence probability is as follows: ; in, The total number of preset risk factors; For risk factor indexing; For risk exist The probability distribution function of the occurrence at time t; For risk The impact intensity vector; Background Gaussian white noise that follows a normal distribution; for The intensity of the composite disturbance injected at any given time.

[0122] Step d2: Based on the calculation results of the composite injection model, obtain the external disturbance input at each unified time step and calculate the composite disturbance intensity.

[0123] The results of the composite injection model refer to the quantitative data on the intensity and probability of external disturbances calculated using the composite injection model. Their purpose is to provide specific, stochastic disturbance parameters for subsequent simulations. For example, the results could be a series of specific disturbance events predicted or generated by the model at different time steps, such as the intensity values ​​of demand surges or transportation delays. A unified time step refers to discretizing the continuous time stream into a series of equal-length time intervals during the simulation. Its purpose is to ensure that all simulation events and data updates occur within a synchronous and consistent time frame, facilitating data alignment and processing. For example, a unified time step could be set to 1 hour, 1 day, or 1 week. The external disturbance input refers to the specific numerical value generated by the composite injection model within each unified time step, representing the impact of external environmental uncertainties. Its purpose is to serve as input to the discrete event simulation engine, simulating random events that affect the operation of the supply chain in the real world. For example, the external disturbance input could be the percentage change in market demand within a certain time step. Composite disturbance intensity refers to the degree of impact on the overall operation of a system, quantified by comprehensively considering multiple external disturbance inputs. Its function is to provide a unified index to measure the overall uncertainty level faced by the system within the current time step. For example, different types of external disturbance inputs can be integrated into a single composite disturbance intensity value through methods such as weighted averaging, principal component analysis, or fuzzy comprehensive evaluation.

[0124] Step d3: Use the calculated composite disturbance intensity and the received preliminary scheduling scheme to drive the discrete event simulation engine.

[0125] The discrete event simulation engine is an event-driven simulation tool used to simulate the evolution of a system over time. Its function is to simulate the behavior and interactions of various entities in the supply chain, such as production lines, transport vehicles, and warehouses, based on preset rules and input parameters, thereby evaluating the performance of scheduling schemes under different scenarios. For example, the engine can be a customized model built on simulation software, or a self-developed simulation framework capable of handling event queues and state updates. Driving the discrete event simulation engine means taking the composite disturbance intensity and the preliminary scheduling scheme as input to start and control the simulation engine's operation. Its function is to ensure that the simulation process realistically reflects the execution effect of the scheduling scheme under uncertain environments. For example, the composite disturbance intensity and preliminary scheduling scheme parameters can be passed to the simulation engine through an API interface to trigger the start of the simulation.

[0126] Step d4: In each independent random simulation cycle, the boundary conditions of the preliminary scheduling scheme during execution are updated in real time based on the composite disturbance intensity value at the current moment. The energy production gap value, transportation delay time value, and inventory overflow value under the corresponding cycle batch are extracted to calculate the value of the whole chain loss function.

[0127] Independent stochastic simulation loops refer to repeatedly executing the simulation process multiple times during dynamic simulation, each loop using different stochastic disturbance scenarios. Their purpose is to cover various uncertainties through multiple simulations, thereby obtaining more statistically significant simulation results. For example, hundreds or even thousands of independent Monte Carlo simulation loops can be performed. Real-time updates of the initial scheduling scheme's boundary conditions during execution refer to dynamically adjusting the constraints or environmental parameters faced by the scheduling scheme based on the current intensity of the composite disturbance. Their purpose is to enable the simulation to demonstrate the adaptability of the scheduling scheme to unexpected situations in the real world. For example, when the intensity of the composite disturbance indicates that extreme weather has caused transportation disruptions, the availability of transportation routes can be updated in real time. Energy production gap values, transportation delay time values, and inventory overflow values ​​are key performance indicators for measuring the efficiency and risk of the supply chain. Their purpose is to quantify the performance of the scheduling scheme under specific disturbance scenarios, providing basic data for calculating the loss function of the entire chain. For example, the energy production gap value can refer to the difference between actual production and planned production. The total supply chain loss function is a comprehensive indicator used to quantify the total loss caused by poor execution of scheduling plans or external disturbances across the entire supply chain under specific disturbance scenarios. Its purpose is to provide a single, comparable value to assess the merits and risk levels of scheduling plans. For example, the total supply chain loss function can be calculated from indicators such as energy production gaps, transportation delays, and inventory overflows through weighted summation, penalty functions, or cost models.

[0128] Step d5 involves performing dynamic simulation and risk prediction operations on the preliminary scheduling scheme based on the numerical value of the full-chain loss function.

[0129] Among them, dynamic simulation and risk prediction operation refers to using the full-chain loss function value as the core basis to conduct a comprehensive performance evaluation and potential risk identification of the preliminary scheduling scheme. Its role is to deeply analyze the performance of the scheduling scheme under various uncertainty scenarios and identify its weak links and potential risk points. For example, by analyzing the statistical characteristics such as the distribution, mean, and variance of the loss function value, the average performance and stability of the scheduling scheme can be evaluated.

[0130] Step d6: Read all full-chain loss function numerical distribution sequences from the device's simulation log file, calculate the quantiles of the full-chain loss function numerical distribution sequences in conjunction with the set confidence interval level, obtain the value at risk threshold, and calculate the tail extreme risk expected value based on the selected subset of data that exceeds the value at risk threshold.

[0131] The simulation log file stores all full-chain loss function values ​​and other relevant simulation data generated by the discrete event simulation engine in multiple simulation cycles. Its purpose is to provide raw data for subsequent statistical analysis and risk assessment. For example, the log file can be a CSV file, a database record, or a specific data structure. The full-chain loss function value distribution sequence refers to the set of all full-chain loss function values ​​extracted from the simulation log file, arranged according to their numerical magnitude or chronological order. Its purpose is to show the overall distribution characteristics of the loss function values, providing a statistical basis for risk quantification. For example, a histogram of the loss function values ​​can be plotted. The confidence interval level is a statistical concept representing the range within which the loss function value may fall at a given confidence level. Its purpose is to define the risk boundary and help determine which loss values ​​are considered "high-risk" or "extremely risky." For example, the confidence interval level can be set to 95% or 99%. Quantile calculation is a statistical method used to determine the value corresponding to a certain percentage of data points in a dataset. Its purpose is to find the corresponding risk value threshold from the loss function value distribution sequence based on the set confidence interval level. For example, if the confidence interval level is 99%, then the 99th percentile is calculated. The Value at Risk (VaR) threshold is a critical value representing the upper limit to which system losses will not exceed a given confidence level. Its function is to judge the level of risk; losses exceeding this threshold are considered extreme risk events. For example, if the calculated VaR threshold is 1 million yuan, then any loss exceeding 1 million yuan is considered high risk. The tail extreme risk expectation refers to the average value of the loss function across all subsets of data exceeding the VaR threshold. Its function is to quantify the average level of loss the system might face when extreme risk events occur, thus providing a more comprehensive assessment of extreme situations. In one example, the Monte Carlo method is used for 10,000 random simulation cycles, and an independent random number generator seed is assigned to each simulation cycle to ensure data independence between simulation executions.

[0132] In each independent stochastic simulation cycle, the discrete event simulation engine updates the boundary conditions of the initial scheduling scheme in real time based on the current composite disturbance intensity value. At the end of the simulation process, the energy production gap, transportation delay time, and inventory overflow value for the corresponding cycle batch are extracted. Based on these values, the total chain loss function value for that cycle is calculated and recorded in the device's simulation log file.

[0133] After all the set random simulation cycles have been completed, the numerical distribution sequence of all full-chain loss functions in the device's simulation log file is read. Quantiles are calculated on the numerical distribution sequence of the full-chain loss functions, using the set confidence interval level, to obtain the Value at Risk (VaR) threshold. Based on the subset of data exceeding the VaR threshold, a quantitative indicator of systemic risk under extreme conditions is calculated. The specific formula used for assessing systemic risk based on conditional VaR is as follows: ; in, The water level is set within the confidence interval. This is the global scheduling instruction matrix over all time steps; It is a set of composite disturbance intensities; In order to schedule instructions in the global instruction matrix Combined perturbation intensity set The full-chain loss function under the given conditions; For the corresponding The risk value threshold; For mathematical expectation operators; This represents the expected value of extreme tail risk.

[0134] Step d7 involves structurally integrating the data set with completed timestamp alignment to generate a full-chain simulation and evaluation matrix covering production nodes, transportation networks, warehousing nodes, and consumer terminals.

[0135] The timestamp-aligned dataset refers to data from different sources and of different types that are timestamped during the simulation. These timestamps have been processed to ensure they are synchronized and consistent in time. Its purpose is to guarantee the temporal correlation between different data, providing an accurate foundation for subsequent integration and analysis. For example, a unified time synchronization protocol can be used. Structured integration refers to organizing and merging timestamp-aligned data from different stages and modalities according to a pre-defined structure and logic. Its purpose is to form a unified and complete data view, facilitating a comprehensive evaluation of the entire industry chain. For example, relational databases, graph databases, or multidimensional arrays can be used. The full-chain simulation evaluation matrix is ​​a multidimensional data structure that contains performance indicators and loss values ​​of each stage of the industry chain under different disturbance scenarios during the simulation. Its purpose is to provide a comprehensive and detailed evaluation result, showing the impact of the scheduling scheme on the entire industry chain. For example, this matrix can include information such as the output of each production node and the delay of each transportation path.

[0136] Step d8 involves packaging and encapsulating the generated full-chain simulation and evaluation matrix, and encoding it using a standard data exchange format to generate a full-chain simulation and evaluation report.

[0137] Packaging and encapsulation refer to organizing the entire chain simulation evaluation matrix and other related simulation results, statistical analysis data, etc., into a logical unit. Its purpose is to facilitate data transmission, storage, and management. For example, it can be packaged into a compressed file. Standard data exchange formats refer to industry-accepted data formats used for data transmission and sharing between different systems. Their purpose is to ensure that the generated evaluation report has good interoperability and readability, facilitating parsing and utilization by other systems or users. The entire chain simulation evaluation report is the final document or data file containing the performance evaluation and risk prediction results of the preliminary scheduling scheme under various disturbance scenarios. Its purpose is to provide decision-makers with comprehensive information to support the optimization and decision-making of the scheduling scheme. For example, the report may include loss function distribution plots, value-at-risk thresholds, and expected values ​​of extreme risks.

[0138] It is understood that, through the above-described technical solution in this embodiment, this application can accurately simulate the random changes in the external environment by setting a dynamic disturbance variable generation program and constructing a composite injection model of random disturbance intensity and occurrence probability, thus overcoming the simulation distortion problem caused by the failure to effectively simulate uncertain factors in the prior art. Based on the calculation results of the composite injection model, the external disturbance input quantity at each unified time step is obtained and the composite disturbance intensity is calculated, ensuring that the disturbance is quantified in the time dimension, providing a dynamic and consistent input benchmark for subsequent simulations. The calculated composite disturbance intensity and the received preliminary scheduling scheme are used to drive the discrete event simulation engine, making the simulation process more closely resemble the actual operating scenario and improving the reliability of the simulation results. In each independent random simulation cycle, the boundary conditions of the preliminary scheduling scheme during execution are updated in real time according to the current composite disturbance intensity value, dynamically adjusting the boundary to adapt to external changes, avoiding the simulation from deviating from the real situation due to fixed boundaries. By extracting the energy production gap value, transportation delay time value, and inventory overflow value and calculating the full-chain loss function value, this application can comprehensively quantify key performance indicators, providing a data foundation for risk analysis. Dynamic simulation and risk prediction are performed based on the numerical values ​​of the entire chain loss function, ensuring that the risk identification process is based on actual performance data and improving the accuracy of the prediction. Furthermore, by reading the distribution sequence of all the entire chain loss function values ​​from the device's simulation log file and combining it with the set confidence interval level, quantile calculations are performed to obtain the value-at-risk (VAT) threshold. Based on the selected subset of data exceeding the VAT threshold, the expected value of tail-end extreme risks is calculated. This application can quantify the risk level through statistical analysis, avoid subjective judgment errors, and assess potential risks for extreme events, enhancing the system's ability to respond to sudden disturbances. Finally, the timestamp-aligned data set is structurally integrated to generate a full-chain simulation and evaluation matrix covering production nodes, transportation networks, warehousing nodes, and consumer terminals. This matrix is ​​packaged and encoded using a standard data exchange format to generate a full-chain simulation and evaluation report, providing comprehensive and standardized evaluation results. This greatly facilitates subsequent scheduling scheme optimization decisions, thereby improving the robustness and decision quality of the entire industry chain intelligent scheduling method.

[0139] In some optional embodiments, this application further proposes extracting a comprehensive performance index sequence from the full-chain simulation and evaluation report, and performing scheduling scheme optimization based on the comprehensive performance index sequence to obtain a set of scheduling decision instructions, specifically including: Step e1: Read the tensor data in the memory cache, construct a multi-agent game and multi-objective joint optimization model based on the tensor data, and set up multiple parallel optimization sub-objectives in the multi-objective joint optimization model.

[0140] Specifically, reading tensor data from the memory cache aims to acquire real-time or near-real-time data, stored in tensor form for efficient processing. This process can be performed via direct memory access (DMI) from the cache or main memory, ensuring low latency in data transmission. A multi-agent game theory and multi-objective joint optimization model is constructed based on tensor data. The goal is to utilize fused multimodal data to establish a mathematical model that simultaneously considers the interactions between multiple decision-making agents and multiple optimization objectives. This model can employ non-cooperative game theory to model the competitive or cooperative relationships between agents, while combining multi-objective programming methods to handle multiple optimization objectives. Multiple parallel sub-objectives are established in the multi-objective joint optimization model to clearly define multiple objectives that need to be optimized simultaneously. These objectives are often conflicting or involve trade-offs, such as minimizing total cost, maximizing production efficiency, minimizing transportation time, minimizing inventory backlog, and maximizing system robustness. These sub-objectives can include cost minimization, efficiency maximization, and risk minimization; alternatively, they can include resource utilization maximization, environmental impact minimization, and customer satisfaction maximization.

[0141] In one example, the specific formula used to construct the multi-agent game and multi-objective joint optimization model is as follows: ; in, This is the global scheduling instruction matrix; The joint optimization objective function that needs to be minimized globally; In order to schedule instructions in the global instruction matrix The total cost across the entire supply chain; In order to schedule instructions in the global instruction matrix Overall efficiency; , , Pareto normalized weights; Water level for the set confidence interval The corresponding expected value of extreme tail risk; This is the absolute value penalty term for imbalance between supply and demand nodes.

[0142] Step e2 assigns corresponding Pareto normalized weights to multiple parallel optimization sub-objectives and performs linear weighted calculation on the multiple objectives to generate a joint optimization objective function that needs to be minimized globally.

[0143] Specifically, Pareto normalized weights are assigned to multiple parallel optimization sub-objectives. This aims to assign importance weights to different optimization sub-objectives and, through Pareto normalization, enable effective comparison and weighting of objectives with different dimensions and ranges. The initial weights can be determined using expert scoring or the analytic hierarchy process (AHP), and then Pareto normalization is used to eliminate the influence of dimensions. Linear weighted calculations are then performed on the multiple objectives to generate a joint optimization objective function that needs to be minimized globally. This aims to combine the normalized and weighted sub-objective functions into a single, optimizable objective function, typically in the form of a weighted sum for easier solution. This can be achieved by multiplying all normalized sub-objective functions by their corresponding weights and then summing them to form a total weighted objective function, which is usually set to be minimized.

[0144] Step e3: Use deep reinforcement learning algorithm to solve the multi-agent game and multi-objective joint optimization model, and calculate the output global scheduling instruction matrix value.

[0145] Specifically, deep reinforcement learning algorithms are employed to solve multi-agent game and multi-objective joint optimization models. The aim is to leverage the powerful perception and decision-making capabilities of deep reinforcement learning to learn optimal scheduling strategies in dynamic environments to address multi-agent game dynamics and multi-objective conflicts. This algorithm can utilize a multi-agent deep Q-network, where each agent represents a decision-making entity in the industry chain, learning the optimal strategy through interaction with the environment. The algorithm calculates and outputs a global scheduling instruction matrix, aiming to generate a series of specific, executable scheduling instructions based on the learned strategies. These instructions are represented in matrix form, containing detailed operations for each link in the industry chain. This global scheduling instruction matrix can include production plans for each production node, route planning for each transportation unit, and inventory transfer instructions for each warehousing node.

[0146] In one example, a parameterized policy network based on a deep neural network is constructed to establish the mapping relationship between the current state of the system and the corresponding action output. The global state matrix of the system is set as the input feature matrix of the parameterized policy network, and the global scheduling instruction matrix is ​​set as the action output tensor of the parameterized policy network.

[0147] Calculate the corresponding action probability distribution sequence based on the global scheduling instruction matrix output by the parameterized policy network under a given system global state matrix. Record the actual system operating state and collect the state access distribution sequence formed by the interaction between the current parameterized policy network and the simulation environment. Calculate the advantage function that evaluates the current action as superior to the average baseline to quantify the payoff difference of the corresponding action.

[0148] An iterative update procedure is performed on the parameterized policy network based on the policy gradient algorithm. By calculating the policy gradient and adjusting the deep neural network parameters along the gradient direction, the global expected cumulative reward of reinforcement learning is maximized. The specific formula used to update the network parameters based on the policy gradient to maximize the expected reward is as follows: ; in, These are parameters for a deep neural network. To enhance the cumulative reward of the overall learning expectation; This is the system's global state matrix; This is the global scheduling instruction matrix; For parameterized policy networks, that is... Lower output The probability of; State access distribution; To evaluate the advantage function of the current action over the average baseline; For mathematical expectation operators; For deep neural network parameters Operators for finding gradients.

[0149] Step e4: Based on the system global state matrix of the current unified time step, input the system global state matrix into the parameterized policy network to obtain the scheduling decision instruction set.

[0150] Specifically, the system global state matrix based on the current unified time step aims to obtain comprehensive state information of the entire industrial chain system at the current moment. Organized in matrix form, it reflects the system's operational status. This global state matrix can include the real-time operating status of each device, inventory levels at each stage, congestion conditions on each transportation route, market demand forecasts, etc. Inputting the system global state matrix into a parameterized strategy network aims to use the current system state as input, and through a pre-trained neural network with adjustable parameters, predict or generate optimal scheduling decisions. This strategy network can be a deep neural network, such as a convolutional neural network or a recurrent neural network, whose parameters are obtained through deep reinforcement learning. The resulting set of scheduling decision instructions aims to have the strategy network output a series of specific, executable scheduling decisions based on the input state matrix. These decisions constitute the final set of scheduling instructions. This set of scheduling decision instructions can include operation instructions for production equipment, route optimization instructions for logistics vehicles, inventory adjustment instructions, and personnel scheduling instructions, etc.

[0151] It is understood that, through the above technical solution in this embodiment, this application reads tensor data from the memory cache and constructs a multi-agent game and multi-objective joint optimization model based on this data. Multiple parallel optimization sub-objectives are established, ensuring that the optimization process can utilize the latest system data in real time, while independently handling the game relationships and multi-objective requirements of different agents, avoiding mutual interference between optimization objectives. By assigning corresponding Pareto normalized weights to multiple parallel optimization sub-objectives and performing linear weighted calculations on the multiple objectives, a global joint optimization objective function to be minimized is generated, effectively balancing the priority conflicts between different optimization objectives, transforming the complex multi-objective problem into a single optimizable function, and simplifying the solution process. A deep reinforcement learning algorithm is used to solve the multi-agent game and multi-objective joint optimization model, calculating and outputting a global scheduling instruction matrix value. Utilizing the dynamic learning capability of reinforcement learning to adapt to environmental changes solves the problem that static optimization methods cannot cope with sudden disturbances, realizing the dynamic generation of scheduling instructions. Based on the system's global state matrix at the current unified time step, the matrix is ​​input into a parameterized policy network to obtain a set of scheduling decision instructions. This ensures that decisions are based on real-time updated system states, and the parameterized policy enhances the model's adaptability and robustness, avoiding decision lag or failure. Combining multi-agent game theory, multi-objective joint optimization, and deep reinforcement learning, the optimization of the scheduling scheme comprehensively considers the conflicts and collaborations among entities in the industry chain, while also taking into account multiple performance indicators. Furthermore, it can adjust the policy in real time according to dynamically changing system states, thereby improving the comprehensiveness, dynamic adjustment capability, and system robustness of scheduling decisions. This addresses the problem that existing methods fail to effectively handle multi-agent game theory and multi-objective conflicts during optimization, leading to incomplete optimization results, insufficient dynamic adjustment, and poor system robustness.

[0152] In some embodiments, this application further proposes a preliminary scheduling scheme for: collecting the local scheduling strategies selected by all agents and their corresponding local net utility values; performing multi-dimensional conflict detection on all local scheduling strategies within the spatial topology framework of the entire industry chain; for resource overload nodes and congested road segments where conflicts are detected, performing resource degradation allocation operations based on the descending order of the local net utility values ​​of each agent; integrating the various local scheduling strategies after conflict detection and resource degradation allocation operations, and outputting a preliminary scheduling scheme based on the integrated global strategy data, including detailed task allocation data for each node, a list of equipment and material resource allocations, and a cross-node transportation path planning matrix.

[0153] Specifically, when generating an initial scheduling scheme, it is first necessary to collect the local scheduling policies selected by all agents and their corresponding local net utility values. This aims to ensure that before making global scheduling decisions, the system can comprehensively obtain the preliminary decisions made by each agent based on its own local information and objectives, as well as quantitative indicators for evaluating the merits of these decisions. For example, through multi-agent communication protocols, such as message queues or shared memory, a central coordinator or a master agent can periodically request and receive the current local scheduling policies and corresponding local net utility values ​​from all participating agents. Alternatively, after completing their local decisions, agents can encapsulate their local scheduling policies and local net utility values ​​into standard data packets and upload them to a global state database or distributed ledger via a pre-defined API interface for other agents or the coordination module to query and summarize.

[0154] All local scheduling strategies are subjected to multi-dimensional conflict detection within the spatial topology framework of the entire industry chain. This aims to identify resource competition or logical contradictions that may result from independent decisions by different agents. Detection within the spatial topology framework of the entire industry chain means considering not only individual resource points but also potential conflicts throughout the entire logistics, information, and capital flow network, such as path intersections, equipment contention, and insufficient inventory. For example, a conflict detection engine can be built that receives all local scheduling strategies and simulates their operation based on the spatial topology framework of the entire industry chain. By traversing all resource nodes and connecting paths, it checks whether multiple agents simultaneously request the same resource, or whether the requested resource quantity exceeds their carrying capacity, and whether multiple paths occupy the same road segment simultaneously, causing congestion. Another approach is to use a rule-based expert system or a constraint satisfaction problem solver. Each local scheduling strategy is transformed into a set of resource occupancy constraints, and then all constraint sets are input into the solver. The solver searches within the spatial topology framework of the entire industry chain for possible combinations of constraints that cannot be simultaneously satisfied, thereby identifying conflict points.

[0155] For resource-overloaded nodes and congested road segments where conflicts are detected, resource degradation allocation is performed based on the descending order of the agents' local net utility values. When conflicts occur, a mechanism is needed to resolve resource competition. Resource degradation allocation is a prioritization strategy that maximizes overall benefits in resolving local conflicts by prioritizing the needs of agents that bring higher local net utility values. For example, for each conflict point, the system first identifies all agents involved in the conflict and their corresponding local scheduling strategies. Then, it extracts the local net utility values ​​contributed by these agents to the conflict point and sorts them in descending order. Resource degradation allocation can manifest as follows: resource requests from higher-ranked agents are prioritized; resource requests from lower-ranked agents may be partially reduced, delayed, or rescheduled until the conflict is resolved. Alternatively, an auction-based or negotiation-based mechanism can be used. When a conflict is detected, the system can initiate a virtual auction where agents "bid" for conflicting resources based on their local net utility values, with the highest bidder winning the resources. Alternatively, agents can negotiate in multiple rounds based on their respective local net utility values ​​through a pre-defined negotiation protocol until a consensus on resource allocation is reached, with agents having higher net utility values ​​having greater weight in the negotiations.

[0156] Finally, the system integrates the various local scheduling strategies after conflict detection and resource degradation allocation operations. Based on the integrated global strategy data, it outputs a preliminary scheduling scheme containing detailed task allocation data for each node, a list of equipment and material resources, and a cross-node transportation path planning matrix. The aim is to ensure that all local strategies are coordinated and consistent after conflict resolution, forming a unified scheduling scheme without internal contradictions, and transforming it into executable, detailed scheduling instructions. For example, a global scheduling strategy database can be established, updating each agent's adjusted local scheduling strategy to this database. The system iterates through the strategies of all agents to ensure that all resource allocation, task order, path selection, and other information are logically coherent and conflict-free. Simultaneously, the system performs structured processing on the integrated global strategy data, generating different types of output files or data structures. For example, the detailed task allocation data can be a table listing the tasks, quantities, and priorities that each production node needs to execute within a specific time period; the equipment and material resource allocation list can be a list detailing which equipment is occupied by which task at what time, as well as the flow and quantity of materials; the cross-node transportation path planning matrix can be a graph or table indicating the transportation routes and methods that different goods should follow at different time periods.

[0157] In one example, the specific formula used for the agent's local net utility is as follows: ; in, Index for the current agent; For intelligent agents Local scheduling strategy; The expected return from strategy execution; For operating costs; The overflow amount due to violation of local constraints; To constrain the penalty coefficient; This represents the local net utility value.

[0158] It is understood that, through the above-described technical solution in this embodiment, this application can effectively solve the problem of low overall efficiency and uneven resource allocation caused by resource conflicts in the local scheduling strategies independently generated by each agent, such as resource-overloaded nodes or congested road sections. By collecting the local scheduling strategies and net utility values ​​of all agents and performing multi-dimensional conflict detection within the spatial topology framework of the entire industry chain, resource overload and path congestion can be accurately identified. Subsequently, resource degradation allocation is performed according to the descending order of the agents' local net utility values, prioritizing high-value needs, thereby effectively resolving conflicts. Finally, these conflict-resolved local strategies are integrated to generate a coordinated and conflict-free preliminary scheduling scheme. This ensures that the preliminary scheduling scheme considers global resource constraints and the coordination between agents during generation, avoiding resource waste and efficiency loss caused by local conflicts. By prioritizing the needs of high net utility values, resource utilization efficiency is optimized, and the fairness and effectiveness of the overall scheduling scheme are improved. The generated preliminary scheduling scheme is detailed and executable, providing high-quality input for subsequent dynamic simulation and risk prediction, thereby improving the robustness and practicality of the intelligent scheduling method for the entire cross-platform multimodal industrial chain.

[0159] This application provides a cross-platform, multimodal intelligent scheduling device for the industrial chain, such as... Figure 4 As shown, the device includes: The data fusion unit 401 is used to perform data fusion processing on cross-platform industry chain data to obtain multimodal fused data, and to perform semantic modeling processing on the multimodal fused data to obtain a unified semantic description network for the entire industry chain.

[0160] The twin modeling unit 402 is used to construct a multi-layer digital twin model of the entire industry chain based on multi-modal fusion data and a unified semantic description network of the entire industry chain, and to obtain a dynamic twin environment based on the output of the multi-layer digital twin model of the entire industry chain. The agent scheduling unit 403 is used to trigger multiple agents to perform distributed cooperative scheduling operations based on a dynamic twin environment and generate a preliminary scheduling scheme. The simulation and deduction unit 404 is used to perform dynamic simulation and risk prediction on the preliminary scheduling scheme by using the discrete event simulation engine inside the driving device, and generate a full-chain deduction and evaluation report. The decision optimization unit 405 is used to extract the comprehensive performance index sequence from the full-chain simulation and evaluation report, and based on the comprehensive performance index sequence, to perform scheduling scheme optimization and obtain a set of scheduling decision instructions.

[0161] In this embodiment, the device is internally configured with an intelligent scheduling subsystem. This subsystem performs data fusion, twin modeling, agent scheduling, simulation inference, and decision optimization. The intelligent scheduling subsystem includes a discrete event simulation engine, a simulation log file, and a scheduling instruction output interface. The discrete event simulation engine performs dynamic simulation and risk prediction under the drive of the initial scheduling scheme. The simulation log file records the loss function numerical distribution sequence, disturbance input, performance indicators, and risk assessment results generated during the simulation process.

[0162] Specifically, the data fusion unit is configured to perform unified access, data governance, multimodal data fusion, and semantic modeling operations for cross-platform multimodal industry chain data. The multi-source data access mechanism adopts a unified data interface standard, allowing access to the entire industry chain data through application programming interfaces (APIs), message queues, real-time streaming data, and batch processing data. The acquired industry chain data can cover production data, transportation data, inventory data, market demand data, scheduling instruction data, and simulation data. Production data includes coal mine operation data and power plant operation data. Transportation data includes railway operation data, port operation data, and shipping operation data. The data fusion unit performs data cleaning, data standardization, metadata management, and data quality assessment operations. Structured data, unstructured data, and time-series data are uniformly mapped to an entity-relationship attribute data model. Structured data is obtained based on relational database storage, unstructured data includes text and image data, and time-series data is collected from field sensor networks. Semantic modeling operations are based on knowledge graph models and ontology models to achieve unified semantic description and association mapping of entity data across the entire industry chain.

[0163] The twin modeling unit constructs a multi-layered digital twin model of the entire industry chain based on fused cross-platform multimodal industry chain data. Modeling objects can cover coal mine entities, port entities, power plant entities, railway transportation networks, shipping routes, inventory nodes, and dispatch centers. The multi-layered digital twin model includes an actual equipment and system layer, a real-time operational data layer, a behavioral model layer, and a dynamic evolution simulation layer. The actual equipment and system layer establishes a virtual mapping topology for the underlying hardware devices; the real-time operational data layer establishes connections with historical features and updates node states in real time. The behavioral model layer describes the state transition laws of each node based on system dynamics; and the dynamic evolution simulation layer drives the evolution of the entire system's operational state. The twin modeling unit employs discrete event simulation, system dynamics, complex network modeling, and multi-agent modeling methods. The multi-agent modeling mechanism transforms industry chain nodes into corresponding agents, covering production agents, transportation agents, warehousing agents, consumption agents, and dispatch agents. Each agent possesses independent state variables, a set of behavioral rules, and decision execution logic. Combining geographic information systems and 3D modeling technology, the twin modeling unit generates a dynamic twin environment using geographic location coordinates, path network topology, and transportation path coordinates.

[0164] The agent scheduling unit executes multi-agent collaborative scheduling operations based on a multi-layered digital twin model across the entire industry chain. The multi-agent collaboration mechanism utilizes distributed decision-making logic, local autonomy rules, and global coordination strategies to perform resource scheduling. Communication between agents involves data exchange via message queues, shared state memory, and event-triggered interfaces. Scheduling strategy outputs include task allocation instructions, resource allocation instructions, and path planning instructions. The agent scheduling unit employs a discrete event-driven engine and an exception-triggered interrupt mechanism to respond to dynamic event inputs and trigger corresponding local scheduling decisions and global state coordination.

[0165] The simulation and extrapolation unit performs dynamic simulation and risk prediction operations on the generated preliminary scheduling plan. It runs a discrete event simulation model, supporting multi-timescale simulation evolution with hourly, daily, and monthly units. A disturbance injection mechanism adds dynamic disturbance variables to the dynamic simulation environment, encompassing weather change data, transportation disruption events, and demand fluctuations. The simulation process of the unit sequentially executes scenario construction, parameter setting, simulation execution, and result output. In one example, the unit can apply the Monte Carlo method to perform multiple discrete event stochastic simulations, quantitatively assessing the robustness indicators and systemic risk probability of the supply chain system. The full-chain simulation covers all business links in production, transportation, sales, storage, and use, outputting a full-chain simulation evaluation report.

[0166] The decision optimization unit extracts comprehensive system performance indicators based on the full-chain simulation and evaluation report, and automatically optimizes the scheduling scheme and outputs the optimal decision command. The comprehensive system performance indicators cover system operating cost indicators, resource utilization efficiency indicators, and expected risk indicators. The multi-objective optimization objectives of the decision optimization unit are configured as minimizing total cost, maximizing total efficiency, minimizing the expected value of extreme risks, and balancing the total supply and demand of nodes.

[0167] The decision optimization unit's optimization algorithm subunit integrates deep reinforcement learning, genetic algorithms, particle swarm optimization, dynamic programming, and robust optimization algorithms. A multi-objective optimization mechanism constructs a multi-objective joint optimization function and applies a Pareto normalized weight allocation mechanism to calculate the comprehensive optimal solution. The game theory model subunit addresses the collaborative conflict problem between nodes by setting multi-agent interest balance constraints. The self-learning mechanism subunit uses a deep reinforcement learning model to solve for the optimal scheduling strategy, issues optimal scheduling decision instructions to each execution entity, and performs closed-loop iterative updates to the deep neural network parameters based on historical scheduling data fed back from the underlying system.

[0168] In one example, the technical effect reference of this application Figure 5 and Figure 6 .like Figure 5 As shown, the horizontal axis of the graph represents the unified time step (h), ranging up to 70 hours; the vertical axis represents the total cost of the entire chain (yuan), ranging from 70,000 to 95,000. The graph includes two comparison curves and their legends: one is a solid black line with hollow circles, representing the static rule-based scheduling strategy; the other is a dark gray dashed line with hollow squares, representing the scheduling decision of this application. This application provides a specific application embodiment of a cross-platform multimodal intelligent scheduling device for the energy industry chain, verifying the parameter calculation logic during system operation.

[0169] In this application embodiment, the data fusion unit receives cross-platform industry chain data, using coal mine mining node data as an example. Taking 1 as an example, set The time is 1 hour. The data fusion unit acquires the time window offset. Raw data parameters for 2 hours The value is 500t. The data fusion unit calls the time interpolation alignment function. The raw data parameters are processed, and the output value is 510t. The data fusion unit reads the historical average parameters of this coal mine mining node from the local database. The value is 480t, and the standard deviation parameter is read. The value is 20t. The data fusion unit substitutes the acquired raw data parameters, time window offset, historical mean parameters, and standard deviation parameters into the spatiotemporal alignment and normalization calculation formula for multi-source time series data to calculate the processed standard data value at this unified time step. It is 1.5.

[0170] The data fusion unit extracts the processed standard data and sets the total number of data modalities. The data modalities are 3. These include relational structured modalities, textual unstructured modalities, and temporal sequence modalities. The data fusion unit extracts independent feature vectors from the textual unstructured modalities. and its corresponding feature mapping weight matrix The dynamic attention weights of this mode at the current unified time step are calculated. The value is 0.4. The data fusion unit uses the same calculation rules to derive the dynamic attention weights for relational structured modalities. Dynamic attention weights of 0.35 and temporal modalities The value is 0.25. The data fusion unit extracts the dynamic attention weights of the three modalities. and the corresponding transformed feature representation matrix Substituting these values ​​into the multimodal feature fusion calculation formula based on dynamic attention mechanism, a cross-modal fusion feature vector matrix is ​​generated. .

[0171] The decision optimization unit constructs a multi-objective joint optimization model based on the data output by the simulation unit. The decision optimization unit extracts data from a given global scheduling instruction matrix. Total cost of the entire supply chain The value is 150,000 yuan, and the total efficiency value is extracted. The confidence interval level is set to 0.85. The expected value of the tail extreme risk corresponding to the water level in this confidence interval is 95%. The value is 5000. The decision optimization unit reads the absolute value penalty term parameter for the supply and demand node imbalance. The value is 200. The decision optimization unit is the Pareto normalized weight value assigned to the total cost of the entire chain. The Pareto normalized weight is set to 0.5, which is the total efficiency value. The Pareto normalized weight is set to 0.3, which is the value used to assign Pareto normalized weights to the expected value of tail extreme risks. It is 0.2.

[0172] The decision optimization unit will use the above-obtained data , , as well as The specific values ​​and corresponding Pareto normalized weights , , Substituting these values ​​into the formula of the multi-agent game and multi-objective joint optimization model, the numerical value of the joint optimization objective function that needs to be minimized globally is calculated. The value is 76199.745. The decision optimization unit uses this value as a baseline and initiates subsequent deep reinforcement learning algorithm solutions to find a global scheduling instruction matrix that further reduces the value of the joint optimization objective function. .

[0173] For the experimental verification and effect comparison section, please refer to the appendix. Figure 6 The graph shows the number of collaborative iterations on the horizontal axis, ranging from 0 to 1000; and the cumulative expected reward on the vertical axis, ranging from -6000 to 4000. The legend displays two types of data: the light gray thin line represents the original collected values, showing the real noise fluctuations during the algorithm exploration process; the thick black line marked with hollow circles represents the smooth convergence trend, reflecting the core trajectory of the algorithm optimization. This application provides an experimental verification and effect comparison implementation method for a cross-platform multimodal industrial chain intelligent scheduling device, which may include: The decision optimization unit sets benchmark parameters within the training environment of the deep reinforcement learning algorithm. A static rule-based scheduling strategy is set as the comparison group, which relies on a fixed inventory threshold to trigger transfer instructions and does not perform dynamic parameter updates. The decision optimization unit extracts the total number of consecutive sampling observation steps from the twin modeling unit. The actual output state values ​​of the real system when there are 1000 values. and the numerical output of the twin model simulation The decision optimization unit substitutes the above values ​​into the system fidelity evaluation formula and combines it with the error diagonal penalty matrix. The system fidelity index of the digital twin spatiotemporal mapping was calculated. The fidelity index is 0.985. This indicates that the constructed multi-layered digital twin model covering the entire industry chain meets the accuracy requirements for algorithm training.

[0174] Initialize deep neural network parameters in the decision optimization unit The policy gradient algorithm is then executed for iterative updates. In the initial phase, the parameterized policy network randomly outputs a global scheduling instruction matrix. Substituting the values ​​into the advantage function calculation formula, we can evaluate and obtain the advantage function that the current action is superior to the average baseline. It fluctuates around zero. As the number of collaborative iterations increases, the decision optimization unit continues to move along the gradient direction. Adjust the parameters. Record the data and substitute it into the reinforcement learning global expected cumulative reward formula to calculate the value. The value gradually increases from the negative range at the beginning of the iteration. (Combined with...) Figure 6 Specifically, both the light gray thin line representing the original collected values ​​and the thick black line marked with hollow circles representing the smooth convergence trend started from around -5000 when the number of collaborative iterations was 0 and then rapidly increased. When the number of collaborative iterations reached around 500, the upward slope of the global expected cumulative reward value gradually decreased and tended to flatten, at which point the reward value climbed to close to 3000. As iterations continued, the reward difference between two consecutive iterations decreased to below the set convergence threshold parameter of 0.01, and then smoothly converged to around 3000 when the number of collaborative iterations reached 1000, triggering the stopping condition.

[0175] The decision optimization unit extracts the updated deep neural network parameters to generate the optimal scheduling decision instruction set, and applies it to the simulation unit for a 72-hour continuous span simulation test. According to the comprehensive performance index sequence extracted from the full-chain simulation evaluation report, the expected value of the tail extreme risk of the entire system after executing the optimal scheduling decision instruction set decreased by 35% compared to the initial state. Simultaneously, combined with the attached... Figure 5 The data shows that the total cost of the scheduling decision in this application, marked with hollow squares, steadily decreased from an initial value of over 75,000 yuan, eventually being controlled at around 71,000 yuan; while the total cost of the comparison group under the static rule scheduling strategy, marked with hollow circles, showed a continuous fluctuating upward trend with a unified time step, eventually climbing to over 90,000 yuan. The above charts and data comparison results confirm that the decision optimization logic of this application achieves joint optimization of various comprehensive performance indicators.

[0176] Figure 7 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application.

[0177] The following is a detailed reference. Figure 7 The diagram illustrates a structural schematic suitable for implementing the electronic device described in the embodiments of this application. The electronic device may include a processor (e.g., a central processing unit, graphics processor, etc.) 701, which can perform various appropriate actions and processes according to a program stored in read-only memory (ROM) 702 or a program loaded from memory 708 into random access memory (RAM) 703. The RAM 703 also stores various programs and data required for the operation of the electronic device. The processor 701, ROM 702, and RAM 703 are interconnected via a bus 704. An input / output (I / O) interface 705 is also connected to the bus 704.

[0178] Typically, the following devices can be connected to I / O interface 705: input devices 706 including, for example, touchscreens, touchpads, keyboards, mice, cameras, microphones, accelerometers, gyroscopes, etc.; output devices 707 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; memory devices 708 including, for example, magnetic tapes, hard disks, etc.; and communication devices 709. Communication device 709 allows electronic devices to exchange data via wireless or wired communication with other devices. Although Figure 7 Electronic devices with various devices are shown, but it should be understood that it is not required to implement or have all of the devices shown, and more or fewer devices may be implemented or have instead.

[0179] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a non-transitory computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via communication device 709, or installed from memory 708, or installed from ROM 702. When the computer program is executed by processor 701, it performs the functions defined in the cross-platform multimodal supply chain intelligent scheduling method of embodiments of this application.

[0180] Figure 7 The electronic device shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments of this application.

[0181] This application also provides a computer-readable storage medium. The methods described in this application can be implemented in hardware or firmware, or implemented as recordable on a storage medium, or implemented as computer code downloaded over a network and originally stored on a remote storage medium or a non-transitory machine-readable storage medium and then stored on a local storage medium. Thus, the methods described herein can be processed by software stored on a storage medium using a general-purpose computer, a dedicated processor, or programmable or dedicated hardware. The storage medium can be a magnetic disk, optical disk, read-only memory, random access memory, flash memory, hard disk, or solid-state drive, etc.; further, the storage medium can also include combinations of the above types of memory. It is understood that computers, processors, microprocessor controllers, or programmable hardware include storage components capable of storing or receiving software or computer code. When the software or computer code is accessed and executed by the computer, processor, or hardware, the cross-platform multimodal supply chain intelligent scheduling method shown in the above embodiments is implemented.

[0182] A portion of this application can be applied as a computer program product, such as computer program instructions, which, when executed by a computer, can invoke or provide the methods and / or technical solutions according to this application through the operation of the computer. Those skilled in the art will understand that the forms in which computer program instructions exist in a computer-readable medium include, but are not limited to, source files, executable files, installation package files, etc. Correspondingly, the ways in which computer program instructions are executed by a computer include, but are not limited to: the computer directly executing the instructions, or the computer compiling the instructions and then executing the corresponding compiled program, or the computer reading and executing the instructions, or the computer reading and installing the instructions and then executing the corresponding installed program. Here, the computer-readable medium can be any available computer-readable storage medium or communication medium accessible to a computer.

[0183] Although embodiments of this application have been described in conjunction with the accompanying drawings, those skilled in the art can make various modifications and variations without departing from the spirit and scope of this application, and such modifications and variations all fall within the scope defined by the appended claims.

Claims

1. A cross-platform, multimodal intelligent scheduling method for the industrial chain, characterized in that, The method includes: Data fusion processing is performed on cross-platform industry chain data to obtain multimodal fused data, and semantic modeling processing is performed on the multimodal fused data to obtain a unified semantic description network for the entire industry chain; Based on the multimodal fusion data and the unified semantic description network of the entire industry chain, a multi-layer digital twin model of the entire industry chain is constructed, and a dynamic twin environment is obtained based on the output of the multi-layer digital twin model of the entire industry chain. Based on the dynamic twin environment, multiple agents are triggered to perform distributed cooperative scheduling operations to generate a preliminary scheduling scheme. Using the discrete event simulation engine inside the preliminary scheduling scheme driving device, dynamic simulation and risk prediction are performed on the preliminary scheduling scheme to generate a full-chain simulation and evaluation report; Extract the comprehensive performance index sequence from the full-chain simulation and evaluation report, and optimize the scheduling scheme based on the comprehensive performance index sequence to obtain a set of scheduling decision instructions.

2. The method according to claim 1, characterized in that, The process involves data fusion processing of cross-platform industry chain data to obtain multimodal fused data, followed by semantic modeling of the multimodal fused data to obtain a unified semantic description network for the entire industry chain, including: Feature extraction is performed on the cross-platform industry chain data to generate independent feature vectors for each modality; A linear transformation is performed on the independent feature vectors based on the feature mapping weight matrix to generate a feature representation matrix corresponding to each modality; The dynamic attention weight value of the feature representation matrix of each modality in the current system state is calculated based on the dynamic attention mechanism. The dynamic attention weight values ​​are multiplied by the corresponding feature representation matrix, and the product results of all modalities are summed to obtain the multimodal fusion data. Semantic modeling processing is performed on the multimodal fusion data to obtain the unified semantic description network for the entire industry chain.

3. The method according to claim 2, characterized in that, The step of performing semantic modeling processing on the multimodal fusion data to obtain the unified semantic description network for the entire industry chain includes: The multimodal fusion data is semantically mapped using an entity relationship attribute model; wherein, the entity identifiers of the multimodal fusion data are used to match the corresponding industry chain entity nodes. Obtain the attribute list contained in the entity node of the industrial chain, and extract the knowledge graph embedding vector of the entity node of the industrial chain; Based on a preset list of relationship types, obtain the relationship projection matrix associated with the entity nodes in the industrial chain; Multiply the knowledge graph embedding vector by the relation projection matrix to generate an entity projection vector mapped to the corresponding relation space; The semantic relevance score of the entities in the industry chain is calculated based on the generated entity projection vector, and the entities with semantic relevance scores greater than the preset relevance score threshold are connected into edges to obtain the unified semantic description network of the entire industry chain.

4. The method according to claim 1, characterized in that, The construction of a multi-layered digital twin model for the entire industry chain based on the multimodal fusion data and the unified semantic description network for the entire industry chain includes: Extract the structural dimension parameters from the multimodal fusion data, and generate the corresponding three-dimensional mesh model based on the structural dimension parameters; The three-dimensional mesh model is assembled and spliced ​​according to spatial relative positions and connection relationships to form a spatial topology framework for the entire industry chain; Configure corresponding rated operating parameter matrices for each node in the spatial topology framework of the entire industry chain, establish the correspondence between the actual equipment and the spatial topology framework of the entire industry chain, obtain the basic mapping model between the actual equipment and system attributes, and use the basic mapping model as the multi-layer digital twin model.

5. The method according to claim 4, characterized in that, The dynamic twin environment obtained based on the output of the multi-layered digital twin model of the entire industry chain includes: The multimodal fusion data is bound to the corresponding basic mapping model to enable the connection between the real-time running data layer and historical features; Establish causal feedback loops and logistics conservation equations between different entity nodes in the industrial chain, and define a set of differential equations for the state evolution of the industrial chain system; The dynamic twin environment is obtained by overlaying and numerically fusing the three-dimensional mesh model, entity behavior model, geographic information system data, and multi-agent parameters.

6. The method according to claim 1, characterized in that, Based on the dynamic twin environment, the process of triggering distributed cooperative scheduling operations among multiple agents to generate a preliminary scheduling scheme includes: A multi-agent distributed communication network is constructed based on the spatial topology framework in the dynamic twin environment, and a local state sharing information channel is established based on the distributed communication network. Perform a multi-agent cooperative network state consistency update operation. Based on the difference between the current agent's resource belief state and the resource belief state of its neighboring agents, combined with the communication topology adjacency weight and the consistency convergence step size, calculate and update the resource belief state for the next cooperative iteration number. Activate the local decision optimization procedure, calculate the local net utility value corresponding to each local scheduling strategy based on the current updated resource belief state, and select the local scheduling strategy with the largest local net utility value as the initial scheduling scheme for the current agent.

7. The method according to claim 1, characterized in that, The discrete event simulation engine within the device driving the preliminary scheduling scheme is used to perform dynamic simulation and risk prediction on the preliminary scheduling scheme, generating a full-chain deduction and evaluation report, including: A dynamic disturbance variable generation program is set up to simulate random changes in external environmental input, and a composite injection model of random disturbance intensity and occurrence probability is constructed. Based on the calculation results of the composite injection model, the external disturbance input amount at each unified time step is obtained, and the composite disturbance intensity is calculated. The calculated composite disturbance intensity and the received preliminary scheduling scheme are used to drive the discrete event simulation engine; In each independent random simulation cycle, the boundary conditions of the preliminary scheduling scheme during the execution process are updated in real time according to the composite disturbance intensity value at the current moment. The energy production gap value, transportation delay time value and inventory overflow value under the corresponding cycle batch are extracted to calculate the value of the whole chain loss function. Based on the numerical value of the full-chain loss function, dynamic simulation and risk prediction operations are performed on the preliminary scheduling scheme; The simulation log file of the reading device contains the numerical distribution sequence of all full-chain loss functions. Combined with the set confidence interval level, the quantile of the numerical distribution sequence of the full-chain loss functions is calculated to obtain the value at risk threshold. Based on the selected subset of data that exceeds the value at risk threshold, the expected value of the tail extreme risk is calculated. The data sets that have completed timestamp alignment are structured and integrated to generate a full-chain simulation and evaluation matrix covering production nodes, transportation networks, warehousing nodes and consumer terminals; The generated full-chain inference and evaluation matrix is ​​packaged and encapsulated, and the full-chain inference and evaluation report is generated using a standard data exchange format.

8. The method according to claim 1, characterized in that, The step involves extracting a comprehensive performance index sequence from the full-chain simulation and evaluation report, and optimizing the scheduling scheme based on the comprehensive performance index sequence to obtain a set of scheduling decision instructions, including: Read tensor data from the memory cache, construct a multi-agent game and multi-objective joint optimization model based on the tensor data, and set up multiple parallel optimization sub-objectives in the multi-objective joint optimization model; Assign corresponding Pareto normalized weights to the multiple parallel optimization sub-objectives, and perform linear weighted calculation on the multiple objectives to generate a joint optimization objective function that needs to be minimized globally. The multi-agent game and multi-objective joint optimization model is solved by deep reinforcement learning algorithm, and the global scheduling instruction matrix value is calculated and output. Based on the system global state matrix at the current unified time step, the system global state matrix is ​​input into the parameterized policy network to obtain the scheduling decision instruction set.

9. A cross-platform, multimodal intelligent scheduling device for the industrial chain, characterized in that, The device includes: The data fusion unit is used to perform data fusion processing on cross-platform industry chain data to obtain multimodal fused data, and to perform semantic modeling processing on the multimodal fused data to obtain a unified semantic description network for the entire industry chain. The twin modeling unit is used to construct a multi-layer digital twin model of the entire industry chain based on the multi-modal fusion data and the unified semantic description network of the entire industry chain, and to obtain a dynamic twin environment based on the output of the multi-layer digital twin model of the entire industry chain. The agent scheduling unit is used to trigger multiple agents to perform distributed cooperative scheduling operations based on the dynamic twin environment and generate a preliminary scheduling scheme. The simulation and deduction unit is used to perform dynamic simulation and risk prediction on the preliminary scheduling scheme by using the discrete event simulation engine inside the driving device of the preliminary scheduling scheme, and generate a full-chain deduction and evaluation report. The decision optimization unit is used to extract a sequence of comprehensive performance indicators from the full-chain simulation and evaluation report, and to optimize the scheduling scheme based on the sequence of comprehensive performance indicators to obtain a set of scheduling decision instructions.

10. An electronic device, characterized in that, include: The system includes a memory and a processor, which are interconnected. The memory stores computer instructions, and the processor executes the computer instructions to perform the cross-platform multimodal supply chain intelligent scheduling method as described in any one of claims 1 to 8.

Images