Advanced planning and scheduling generation method and system assisted by large model intelligent agent

Abstract

The present application relates to an advanced planning and scheduling generation method and system assisted by a large-model intelligent agent, the method comprising: obtaining the current scheduling intention information expected by the target object from the context information generated by natural language interaction with the target object, wherein the current scheduling intention information includes sub-element parameter items associated with the current APS; based on the current scheduling intention information, retrieving the target mechanism knowledge corresponding to the current scheduling intention information from the mechanism knowledge stored in the mechanism knowledge base corresponding to the preset intelligent agent, each mechanism knowledge being associated with a sub-element parameter item; after obtaining the multi-source production data currently corresponding to the multi-resource constrained manufacturing system, inputting the current scheduling intention information, the multi-source production data and the target mechanism knowledge into the APS construction module corresponding to the intelligent agent, generating the target APS using the retrieval enhancement generation method, and obtaining the advanced planning and scheduling results.

Description

Technical Field

[0001] The present application relates to the technical field of production management of manufacturing enterprises, in particular to an advanced planning and scheduling generation method and system assisted by a large model intelligent agent. Background Art

[0002] With the rapid development of the manufacturing industry, customer order demands are becoming increasingly personalized, characterized by small batches, a wide variety of products, and a high degree of randomness. This places higher demands on production management within manufacturing companies. Advanced Planning and Scheduling (APS), a crucial component of production management for manufacturing companies, is a key component in achieving effective management decisions. The core of APS lies in leveraging dynamic management planning theory and technology to effectively address the decision-making process for multi-task, multi-process production scheduling, taking into account the dynamic demands of customer demand, raw materials, and production equipment, as well as the demands for limited resources such as order changes and production anomalies.

[0003] In related technologies, the APS decision-making method under the traditional production model rarely considers the manufacturing system characteristics of the new production model. Under the new business model of multi-variety and small-batch production, the manufacturing system APS needs to take into account constraints such as order demand, manufacturing resource elements and process requirements; at the same time, the mathematical model of the existing technical solution fails to well describe the multi-resource constraint decision-making scenario, and cannot respond to customized orders in time or adjust decisions according to manufacturing scenario data, and lacks sufficient adaptability; in related technologies, the existing APS technical solutions mostly focus on innovative mathematical models and solution algorithms, the decision-making mechanism is complex, the professional barriers to manufacturing system APS decision-making increase, and there is a lack of multidisciplinary knowledge integration module design. In the design, knowledge is only inherited in a single dimension of a single engineer, and enterprises face challenges in reusing methods / techniques. Faced with dynamically changing market demands and internal resource conditions, they often need to negotiate and try again and again, and cannot quickly utilize existing technologies and solutions. Among related technologies, the APS decision-making of manufacturing systems has problems of information asymmetry and inconsistent goals, resulting in low efficiency of cross-departmental communication and collaboration, high time and labor costs, and easy decision-making delays. Furthermore, there are few APS decision-making paradigms for large-scale model-assisted manufacturing systems, and it is difficult to apply large-scale model technology to manufacturing system APS. There are shortcomings such as output illusions, instability, and high scheduling and execution costs.

[0004] At present, there are problems in APS decision-making of multi-resource constrained manufacturing systems in related technologies, such as insufficient adaptability, few APS decision-making paradigms and poor decision-making results, and no effective solution has been proposed. Summary of the Invention

[0005] The embodiments of the present application provide a large-model intelligent agent-assisted advanced planning and scheduling generation method and system to at least solve the problems in related technologies of insufficient adaptability, few APS decision paradigms and poor decision results in APS decisions of multi-resource constrained manufacturing systems.

[0006] In the first aspect, an embodiment of the present application provides a large-model intelligent agent-assisted advanced plan and scheduling generation method, comprising: obtaining the current scheduling intention information expected by the target object from the context information generated by natural language interaction with the target object, wherein the current scheduling intention information includes sub-element parameter items associated with the current APS, and the sub-element parameter items include at least one of the following: APS sub-objective function, APS sub-constraint parameter, and one APS represents an advanced plan and scheduling; based on the current scheduling intention information, the target mechanism knowledge corresponding to the current scheduling intention information is retrieved from the mechanism knowledge stored in the mechanism knowledge base corresponding to the preset intelligent agent, wherein the mechanism The mechanism knowledge base is generated based on the preset sub-element parameter items and the prompt words are used to stimulate the large model for knowledge expansion. The preset sub-element parameter items are written and generated based on the scheduling scenario characteristic parameters of the multi-resource constrained manufacturing system, and each mechanism knowledge is associated with one sub-element parameter item; after obtaining the multi-source production data currently corresponding to the multi-resource constrained manufacturing system, the current scheduling intention information, the multi-source production data and the target mechanism knowledge are input into the APS construction module corresponding to the intelligent agent, and the target APS is generated using the retrieval enhancement generation method to obtain advanced planning and scheduling results, wherein the multi-source production data is used to characterize the multi-dimensional production status corresponding to the multi-resource constrained manufacturing system.

[0007] In a second aspect, an embodiment of the present application provides an advanced planning and scheduling system, comprising a memory and a processor, wherein the memory stores a computer program, and when the processor executes the computer program, it implements the steps of the large model intelligent agent-assisted advanced planning and scheduling generation method as described in the first aspect.

[0008] In a third aspect, an embodiment of the present application provides a storage medium on which a computer program is stored, which, when executed by a processor, implements the steps of the large-model intelligent agent-assisted advanced planning and scheduling generation method as described in the first aspect above.

[0009] Compared with the related art, the large-model intelligent agent-assisted advanced planning and scheduling generation method and system provided in the embodiment of the present application adopts the context information generated by natural language interaction with the target object to obtain the current scheduling intention information expected by the target object, wherein the current scheduling intention information includes the sub-element parameter items associated with the current APS, and the sub-element parameter items include at least one of the following: APS sub-objective function, APS sub-constraint parameter, and one APS represents an advanced plan and scheduling; based on the current scheduling intention information, the target mechanism knowledge corresponding to the current scheduling intention information is retrieved from the mechanism knowledge stored in the mechanism knowledge base corresponding to the preset intelligent agent. The mechanism knowledge base is generated by using prompt words to stimulate the large model for knowledge expansion based on the preset sub-element parameter items. The preset sub-element parameter items are written and generated based on the scheduling scenario characteristic parameters of the multi-resource constraint manufacturing system. Each mechanism Knowledge is associated with a sub-element parameter item; after obtaining the multi-source production data currently corresponding to the multi-resource constrained manufacturing system, the current planning and scheduling intention information, multi-source production data and target mechanism knowledge are input into the APS construction module corresponding to the intelligent agent, and the target APS is generated using the retrieval enhancement generation method to obtain advanced planning and scheduling results. Multi-source production data is used to characterize the multi-dimensional production status corresponding to the multi-resource constrained manufacturing system, which solves the problems of insufficient adaptability, few APS decision paradigms and poor decision results of the APS decision of the multi-resource constrained manufacturing system in related technologies, and realizes the adaptation to the advanced planning and scheduling needs of the multi-resource constrained manufacturing system, ensures the effectiveness of order production scheduling and production optimization, reduces repetitive negotiation trial and error, quickly utilizes existing technologies and solutions, reduces communication time and labor costs, realizes the automation of the manufacturing system decision-making process, accelerates APS decision generation and improves decision-making effects.

[0010] The details of one or more embodiments of the present application are set forth in the following drawings and description to make other features, objects, and advantages of the present application more readily apparent. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] The drawings described herein are used to provide a further understanding of the present application and constitute a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application and do not constitute an improper limitation on the present application. In the drawings:

[0012] Figure 1 This is a hardware structure block diagram of a terminal of the large model intelligent agent-assisted advanced planning and scheduling generation method according to an embodiment of the present application;

[0013] Figure 2 is a flow chart of a large model agent-assisted advanced planning and scheduling generation method according to an embodiment of the present application;

[0014] Figure 3It is a structural block diagram of an advanced planning and scheduling generation device assisted by a large model intelligent agent according to an embodiment of the present application. DETAILED DESCRIPTION

[0015] In order to make the purpose, technical solutions and advantages of the present application clearer, the present application is described and illustrated below in conjunction with the accompanying drawings and examples. It should be understood that the specific embodiments described herein are merely used to explain the present application and are not intended to limit the present application. Based on the embodiments provided in the present application, all other embodiments obtained by persons of ordinary skill in the art without making any creative work are within the scope of protection of the present application. In addition, it is also understandable that although the efforts made in this development process may be complex and lengthy, for persons of ordinary skill in the art related to the contents disclosed in the present application, some changes such as design, manufacturing or production based on the technical contents disclosed in the present application are merely conventional technical means and should not be understood as meaning that the contents disclosed in the present application are insufficient.

[0016] References to "embodiments" in this application mean that a particular feature, structure, or characteristic described in connection with the embodiment may be included in at least one embodiment of the application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, that the embodiments described in this application may be combined with other embodiments unless there is a conflict.

[0017] Unless otherwise defined, technical or scientific terms used in this application shall have the ordinary meaning as understood by persons of ordinary skill in the art to which this application belongs. The use of "a," "an," "an," "the," and similar expressions in this application does not denote a limitation of quantity and may refer to either the singular or the plural. The terms "comprise," "include," "have," and any variations thereof, as used in this application, are intended to cover non-exclusive inclusions. For example, a process, method, system, product, or device comprising a series of steps or modules (units) is not limited to the listed steps or units but may also include steps or units not listed, or may include other steps or units inherent to the process, method, product, or device. As used in this application, "multiple steps" means two or more steps. "And / or" describes a relationship between associated objects, indicating that three possible relationships exist. For example, "A and / or B" may mean: A exists alone, A and B exist simultaneously, or B exists alone. The terms "first," "second," and "third," etc., as used in this application, simply distinguish similar objects and do not imply a specific ordering of the objects.

[0018] Before describing the embodiments of the present application, the relevant terms involved in the embodiments of the present application are explained as follows:

[0019] A manufacturing system is a collection of workstations and related processes used to assemble products.

[0020] Multiple resource constraints refer to the limitations on the quantity or capacity of resources (such as workers, machines / tools, materials, etc.) in the production system.

[0021] Advanced Planning and Scheduling (APS) refers to the precise arrangement of production resources and planned production times for production and processing tasks, taking into account production resource constraints, through optimization methods, so that production can be completed in a timely manner and resources can be fully utilized. In discrete industries, APS is used to solve the problem of optimal scheduling of multiple processes and multiple resources, while in process industries, APS is used to solve sequential optimization problems.

[0022] A large language model refers to a deep learning model trained using large amounts of text data, which can generate natural language text or understand the meaning of language text.

[0023] An intelligent agent is an agent that can perceive its environment and take actions to achieve specific goals. It can be software, hardware, or a system that has the ability to be autonomous, adaptive, and interactive.

[0024] The following describes the embodiments of the present application with reference to the accompanying drawings:

[0025] The method embodiment provided in this embodiment can be executed in a terminal, a computer or a similar computing device. Taking running on a terminal as an example, Figure 1 This is a hardware block diagram of the terminal of the large model intelligent agent assisted advanced planning and scheduling generation method of the embodiment of the present application. Figure 1 As shown, the terminal may include one or more ( Figure 1 Only one is shown) a processor 102 (the processor 102 may include but is not limited to a microprocessor MCU or a programmable logic device FPGA and other processing devices) and a memory 104 for storing data. Optionally, the terminal may also include a transmission device 106 and an input / output device 108 for communication functions. A person skilled in the art will understand that Figure 1 The structure shown is only for illustration and does not limit the structure of the above terminal. Figure 1 More or fewer components than shown, or with Figure 1 Different configurations shown.

[0026] Memory 104 can be used to store computer programs, such as software programs and modules of application software, such as the computer program corresponding to the large-scale intelligent agent-assisted advanced planning and scheduling generation method in the embodiments of the present invention. Processor 102 executes the computer program stored in memory 104 to execute various functional applications and data processing, thereby implementing the aforementioned method. Memory 104 may include high-speed random access memory and may also include non-volatile memory, such as one or more magnetic storage devices, flash memory, or other non-volatile solid-state memory. In some examples, memory 104 may further include memory remotely located relative to processor 102, and such remote memory may be connected to the terminal via a network. Examples of such networks include, but are not limited to, the Internet, an intranet, a local area network, a mobile communication network, and combinations thereof.

[0027] Transmission device 106 is used to receive or transmit data via a network. A specific example of the aforementioned network may include a wireless network provided by the terminal's communications provider. In one embodiment, transmission device 106 includes a network interface controller (NIC), which can be connected to other network devices via a base station to enable communication with the Internet. In another embodiment, transmission device 106 may be a radio frequency (RF) module, which is used to communicate with the Internet wirelessly.

[0028] This embodiment provides an advanced planning and scheduling generation method assisted by a large model intelligent agent running on the above terminal. Figure 2 This is a flow chart of an advanced planning and scheduling generation method assisted by a large model agent according to an embodiment of the present application, such as Figure 2 As shown, the process includes the following steps:

[0029] Step S201, obtain the current scheduling intention information expected by the target object from the context information generated by natural language interaction with the target object, wherein the current scheduling intention information includes the sub-element parameter items associated with the current APS, and the sub-element parameter items include at least one of the following: APS sub-objective function, APS sub-constraint parameter, and an APS represents an advanced plan and scheduling.

[0030] In this embodiment, the target objects include but are not limited to persons who are expected to obtain the APS solution, for example, a production manager of a multi-resource constrained manufacturing system.

[0031] In this embodiment, the target object conducts multiple rounds of natural language interaction with the large model intelligent agent. The large model clearly identifies the scheduling intention information corresponding to the current scheduling in the corresponding context information management (semantic analysis and recognition). The corresponding scheduling intention information includes the target items, constraint items and selected intelligent optimization algorithms of advanced planning and scheduling. That is, the large model determines the intention vocabulary in the scheduling intention information through semantic recognition, and then matches the relevant target items, constraint items and optimization algorithms corresponding to the corresponding scheduling intention information from the target items, constraint items and intelligent optimization algorithms stored in the corresponding database, that is, matches the corresponding sub-element parameter items.

[0032] Step S202, based on the current scheduling intention information, retrieve the target mechanism knowledge corresponding to the current scheduling intention information from the mechanism knowledge stored in the mechanism knowledge base corresponding to the preset intelligent agent, wherein the mechanism knowledge base is generated based on the preset sub-element parameter items and the prompt words are used to stimulate the large model for knowledge expansion, and the preset sub-element parameter items are written and generated based on the scheduling scenario characteristic parameters of the multi-resource constraint manufacturing system, and each mechanism knowledge is associated with a sub-element parameter item.

[0033] In this embodiment, the large model inputs the current scheduling intention information into the corresponding vector search engine, that is, into the mechanism knowledge base. The vector search engine uses indexing technology to quickly retrieve the mechanism knowledge that is most similar to the current scheduling intention information from the vector knowledge base of the mechanism knowledge base (represented by knowledge items composed of index vectors and knowledge text vectors); during the retrieval process, the similarity between the intention vector corresponding to the current scheduling intention information and the index vector corresponding to each mechanism knowledge in the mechanism knowledge base is calculated to determine which knowledge items corresponding to the mechanism knowledge are most relevant; based on the similarity calculation results, the top K knowledge items ranked by similarity are selected as TOP-K knowledge. These knowledge items cover target items, constraint items and intelligent optimization algorithms that may be related to the current scheduling intention information input by the target object.

[0034] In this embodiment, the corresponding mechanism knowledge base is constructed by first compiling parameter items with target items, constraint items and variant items as sub-elements according to the different scenarios and computing power conditions of the multi-resource constraint manufacturing system to build an initial APS model library; in this embodiment, the initial model library uses target items, constraint items and variant items as sub-elements, and takes minimizing the maximum completion time, minimizing order delays, minimizing the cleaning time and preset time after equipment switching, maximizing capacity utilization, etc. as target items; taking order level sorting constraints, delivery day sorting constraints, product inventory constraints, equipment production line capacity range and order integration or splitting constraints, raw material inventory constraints, equipment occupancy constraints as constraint items; using the above parameters, variables, target items, constraint items and variant items to form the advanced planning and scheduling initial model library; then, using prompt words to stimulate the large model to assist in expanding the APS model library, the corresponding expansion method Including: modifying the names of objectives and constraints, simplifying or complicating mathematical formulas, simulating the expression habits of different users to change the content of notes, and ensuring that the core logic is consistent with the original intention; then, based on the complete model library, guided by the ideas of intelligent optimization algorithms such as simulated annealing algorithm, genetic algorithm, particle swarm algorithm, etc., carry out model instance solving, and create an algorithm library in the form of Python programming code as output; in this embodiment, the mathematical formulas contained in the corresponding objective items, constraint items and variant items in the model library are all presented in LaTeX format with the help of format conversion tools; in this embodiment, after constructing a complete APS model library, the objective items and constraint items in the APS model library are disassembled in the form of questions and answers, and then the knowledge is reconstructed to obtain mechanism knowledge; the corresponding mechanism knowledge is stored in the corresponding database after supervised segmentation, data index construction, feature extraction alignment, structural standardization and other processing.

[0035] In step S203, after obtaining the multi-source production data currently corresponding to the multi-resource constrained manufacturing system, the current planning and scheduling intention information, multi-source production data and target mechanism knowledge are input into the APS construction module corresponding to the intelligent agent, and the target APS is generated using the retrieval enhancement generation method to obtain advanced planning and scheduling results, wherein the multi-source production data is used to characterize the multi-dimensional production status corresponding to the multi-resource constrained manufacturing system.

[0036] In this embodiment, the intelligent agent receives the current scheduling intention information and target mechanism knowledge, integrates and sorts the current scheduling intention information and target mechanism knowledge, forms a context consisting of scheduling intention, index vector and knowledge text vector, and then links the corresponding scene data, and synchronously obtains the multi-source production data corresponding to the multi-resource constraint manufacturing system in the form of data call, and assigns it with the knowledge text vector of the context; in this embodiment, the multi-source production data corresponding to the multi-resource constraint manufacturing system includes but is not limited to the data of the workers, machine status and material status of the multi-resource constraint manufacturing system, specifically including the relevant data corresponding to the warehousing system and the manufacturing execution system, which corresponds to the parameters in the corresponding target items and constraint items in the APS solution. For example, the APS solution is set to a product inventory constraint, and the constraint formula includes "the total amount of material A required to produce the products of the current wave order m≤material A inventory M, all can be scheduled and the material A inventory M can be updated after scheduling", then it is necessary to determine the current inventory quantity of material A in the multi-resource constraint manufacturing system to convert M into a real number.

[0037] In this embodiment, under the action of prompt words composed of instructions and real-numbered context as input data, the large model conducts in-depth analysis and understanding of the real-numbered context. In the learning context, it at least identifies key target items, constraint items and intelligent optimization algorithms, and performs further semantic analysis and logical reasoning on this information to generate the target APS.

[0038] Through the above steps S201 to S203, the current scheduling intention information expected by the target object is obtained from the context information generated by the natural language interaction with the target object, wherein the current scheduling intention information includes the sub-element parameter items associated with the current APS, and the sub-element parameter items include at least one of the following: APS sub-objective function, APS sub-constraint parameter, and one APS represents an advanced plan and scheduling; based on the current scheduling intention information, the target mechanism knowledge corresponding to the current scheduling intention information is retrieved from the mechanism knowledge stored in the mechanism knowledge base corresponding to the preset intelligent agent, wherein the mechanism knowledge base is generated based on the preset sub-element parameter items and the large model is stimulated by prompt words to expand the knowledge, and the preset sub-element parameter items are written and generated based on the scheduling scenario characteristic parameters of the multi-resource constraint manufacturing system, and each mechanism knowledge is associated with a sub-element parameter item. ; After obtaining the multi-source production data currently corresponding to the multi-resource constrained manufacturing system, the current planning and scheduling intention information, multi-source production data and target mechanism knowledge are input into the APS construction module corresponding to the intelligent agent, and the target APS is generated using the retrieval enhancement generation method to obtain advanced planning and scheduling results. Among them, the problems of insufficient adaptability, few APS decision paradigms and poor decision results in the APS decision of the multi-resource constrained manufacturing system in related technologies are solved, and the advanced planning and scheduling requirements of the multi-resource constrained manufacturing system are adapted to ensure the effectiveness of order production scheduling and production optimization, reduce repetitive negotiation trial and error, quickly utilize existing technologies and solutions, reduce communication time and labor costs, realize the automation of the manufacturing system decision process, accelerate APS decision generation and improve decision-making effects. Multi-source production data is used to characterize the multi-dimensional production status corresponding to the multi-resource constrained manufacturing system.

[0039] It should be noted that after generating the target APS, the corresponding code generation module of the intelligent agent is used to generate the corresponding algorithm code based on the method of generating the target APS. The syntactic correctness of the generated algorithm code is also diagnosed by the corresponding code diagnosis module of the intelligent agent, and the algorithm code is corrected based on the diagnosis results. Then, the corrected algorithm code is checked for its problem adaptability, coding logic, code syntax, etc. The inspection results are fed back to the big model, and the final algorithm code is obtained through multiple rounds of interaction. After being transferred to the code execution tool and compiled into natural language, the advanced planning and scheduling decision-making plan for the manufacturing system is obtained.

[0040] It should be further explained that the advanced planning and scheduling generation method of the embodiment of the present application also has the following technical advantages:

[0041] First, it has high adaptability. The mechanism knowledge base and algorithm library of the embodiment of the present application are constructed based on the manufacturing system characteristics of the new production model. The target items, constraint items, and variant items take into account order requirements, manufacturing resource elements, and process requirements. At the same time, by collecting multi-source production data and synchronously updating parameters during modeling, it can adapt to the advanced planning and scheduling requirements of multi-resource constraint manufacturing systems, and ensure the effectiveness of order production scheduling and production optimization; the method of the embodiment of the present application takes into account multi-factor constraints such as order requirements, manufacturing resource elements, and process requirements to construct a model library, and simultaneously creates an intelligent optimization algorithm library. Compared with the existing technology, the method of the embodiment of the present application breaks the traditional modeling and solution methods of advanced planning and scheduling of manufacturing systems. Multiple types of targets, constraints-variant items can be adjusted and combined in time according to decision-making needs, and have high adaptability.

[0042] Second, the degree of reuse of decision-making knowledge / technology is improved. The advanced planning and scheduling generation method of the embodiment of the present application processes the knowledge of models, algorithms, etc. and stores it in a vector database. It has a high degree of integration and convenient knowledge management. It uses the retrieval enhancement generation method to reduce the barriers to the use of knowledge / technology, so that knowledge can break the dilemma of single-dimensional inheritance and reduce repetitive negotiation and trial and error. The advanced planning and scheduling of the manufacturing system can quickly utilize existing technologies and solutions. The method of the embodiment of the present application processes and stores the decision-making knowledge, which can be easily reconstructed. Multidisciplinary knowledge is modularly managed with a high degree of integration. It uses prompt engineering and retrieval enhancement generation technology to enable knowledge to break the traditional single-dimensional inheritance dilemma, improve the application automation of existing technologies and solutions for advanced planning and scheduling of manufacturing systems, and have a high degree of knowledge / technology reuse.

[0043] Third, innovation in cross-departmental communication and collaboration channels: This application uses large-scale intelligent models to assist in the advanced planning and scheduling of multi-resource constrained manufacturing systems, and provides a natural language interactive decision-making method. Model engineers, algorithm engineers, etc. participate in decision-making collaboration through mechanism knowledge management and algorithm code verification. Production managers can apply the proposed advanced planning and scheduling methods and systems to obtain decision-making solutions, share information between manufacturing enterprise departments, and reduce cross-departmental communication time and labor costs.

[0044] Fourth, the decision-making paradigm of the large-scale model-assisted manufacturing system APS has been increased. In a manufacturing environment with transparent and traceable operation processes, the embodiment of the present application starts from the underlying interaction logic and overall operation mechanism of APS, applies prompt engineering and retrieval enhancement generation technology to achieve an effective integration of mechanism intelligence and large-scale model technology, and proposes an APS method and system assisted by a large-scale model intelligent agent, which helps to automate the decision-making process from order receipt to production in the manufacturing system, and provides a large-scale model-assisted professional field decision-making paradigm.

[0045] Fifth, the quality of the output results of the large model in specific scenarios is improved. The embodiment of this application constructs advanced planning and scheduling expertise of the manufacturing system in the form of a model library and an algorithm library, rooted in the scenario, to prompt engineering, retrieval enhancement generation technology to prompt the large model to learn the knowledge of the decision-making mechanism of specific scenarios, realize the construction of the mathematical model of the large model and the generation of algorithm code under the guidance of mechanism wisdom, and improve the quality of the output results of the large model in specific scenarios.

[0046] In some embodiments, based on the current scheduling intention information, the target mechanism knowledge corresponding to the current scheduling intention information is retrieved from the mechanism knowledge stored in the mechanism knowledge base corresponding to the preset intelligent agent, and the steps are as follows:

[0047] Step 21, perform natural language processing and word vector processing on the current scheduling intention information, generate an intention vector corresponding to the current scheduling intention information, and determine the knowledge item corresponding to each mechanism knowledge, wherein the knowledge item includes a first context composed of an index vector and a knowledge text vector, one index vector corresponds to a sub-element parameter item, and the knowledge text vector is used to represent the parameters and variables of the associated index vector.

[0048] In this embodiment, the execution subject model processes the current scheduling intention information into intention vocabulary, and then performs at least word embedding, word mapping, and context combination processing to generate a corresponding intention vector. In this embodiment, the mechanism knowledge exists in the form of knowledge items composed of "QA (question-answer)" question-answer pairs. For example, a mechanism knowledge, Q: the goal is to minimize the maximum completion time, A: the formula is , which states “minimize total production time”.

[0049] Step 22: determine the similarity between the index vectors corresponding to all knowledge items and the intention vector, and select a preset number of index vectors in descending order of similarity to obtain a target index vector.

[0050] In this embodiment, the large model inputs the intention vector into the vector search engine (corresponding to the mechanism knowledge base). The search engine uses indexing technology to quickly retrieve the question-answer pairs most similar to the intention vector from the mechanism knowledge of the mechanism knowledge base. During the retrieval process, the similarity between the intention vector and each "Q (question)" vector of the mechanism knowledge is calculated to determine which knowledge items are most relevant; in this embodiment, the cosine similarity between the intention vector and the index vector (corresponding to the Q in the question-answer pair) is calculated to determine the similarity between the intention vector and the index vector; at the same time, in this embodiment, the top K knowledge items ranked by similarity are selected as TOP-K knowledge, that is, the target mechanism knowledge is obtained, where K is a positive integer greater than or equal to 2.

[0051] Step 23: Use the knowledge entry corresponding to the target index vector as the target mechanism knowledge corresponding to the current itinerary planning intention information.

[0052] Through the above Steps 21 to 23, the most relevant mechanism knowledge is matched from the preset mechanism knowledge, that is, the target items, constraint items, and intelligent optimization algorithms most relevant to the itinerary planning intention information are obtained, providing more valuable background and detailed information for the subsequent large model generation, constructing the learning context, improving the accuracy and adaptability of the generated target APS, and enabling the target object to obtain the most satisfactory decision-making plan for advanced planning and scheduling.

[0053] In some embodiments, natural language processing and word vector processing are performed on the current itinerary planning intention information to generate an intention vector corresponding to the current itinerary planning intention information, which is achieved through the following steps:

[0054] Step 31: Perform natural language processing on the current itinerary planning intention information to generate intention vocabulary corresponding to the current itinerary planning intention information.

[0055] In this embodiment, after the target object proposes the current itinerary planning intention information, the large model performs word segmentation on the natural language corresponding to the current itinerary planning intention information, cuts the corresponding words and sentences into individual words or phrases, removes stop words (including words that frequently appear in the words and sentences but are not very helpful for language understanding, such as "of", "is", "in") from the natural language by comparing with the stop word list, and then converts the remaining words into a machine-readable coding form to obtain the intention vocabulary.

[0056] Step 32: Perform word embedding processing and word mapping processing on the intention vocabulary in sequence to generate corresponding word vectors, where the word vectors are at least used to represent the semantic information of the corresponding intention vocabulary.

[0057] In this embodiment, the large model processes the current itinerary planning intention information into intention vocabulary and converts it into word vectors through word embedding processing. Each intention vocabulary is mapped to a numerical vector, and these vectors capture the semantic information of the intention vocabulary.

[0058] Step 33: Use a preset context encoder to combine all the word vectors into a context semantic vector, where the intention vector includes the context semantic vector.

[0059] In this embodiment, the word vectors generated by processing are combined through the context encoder to generate a semantic representation vector of the current itinerary planning intention information, that is, the intention vector.

[0060] Through the above Steps 31 to 33, the current itinerary planning intention information in natural language is converted into an intention vector, providing a data basis for vector matching of the target mechanism knowledge.

[0061] In some of these embodiments, the current itinerary planning intention information expected by the target object is obtained from the context information generated by the natural language interaction with the target object, and is achieved through the following steps:

[0062] Step 41: Process the context information generated by the target object during the interaction using natural language processing (NLP) to obtain encoded data. The context information includes the target parameters of the current scheduling, constraint requirements, and the intelligent optimization algorithm to be selected. The intelligent optimization algorithm includes one of the following: simulated annealing algorithm, genetic algorithm, particle swarm algorithm. The data processing includes word segmentation, stop word removal, and vocabulary conversion encoding.

[0063] In this embodiment, the large model performs word segmentation on the context information proposed by the target object, cutting the natural language into individual words or phrases. By comparing with the stop word list, stop words (including words that frequently appear in sentences but are not very helpful for language understanding, such as "of", "is", "in") are removed from the natural language. Then, the remaining vocabulary is converted into a machine-readable encoded form to obtain the encoded data.

[0064] Step 42: Use a preset intention recognition network to extract features and recognize intentions from the encoded data, generating multiple candidate intention labels. The candidate intention labels include candidate intention categories and candidate intention category confidence levels. The intention recognition network is an intention recognition model trained using a deep learning neural network based on the attention mechanism and is trained to recognize the intention label corresponding to the input sample Q&A information.

[0065] In this embodiment, the large model uses a deep learning architecture. Through a large number of parameters and a multi-layer neural network structure, it can automatically learn complex feature representations in the text. At the same time, the large model performs deep learning based on the attention mechanism, enabling the large model to focus on the key parts of the input text, assign different weights according to the relevance of the vocabulary to the target intention, and thus better understand the semantics. In this embodiment, the large model calculates the probability distribution of the input natural language belonging to each predefined intention category (corresponding to the candidate intention category), that is, the corresponding candidate intention category confidence level.

[0066] Step 43: Select the target intention label from the multiple candidate intention labels according to the candidate intention category confidence level to obtain the current itinerary planning intention information.

[0067] In some of the optional implementation manners, selecting the target intention label from the multiple candidate intention labels according to the candidate intention category confidence level to obtain the current itinerary planning intention information is achieved through the following steps:

[0068] Step 431 : Select at least one candidate intent label whose candidate intent category confidence is greater than a confidence threshold from among multiple candidate intent labels to obtain a target intent label.

[0069] In this embodiment, at least one candidate intent label with the highest confidence in the candidate intent category is selected to obtain the target intent label.

[0070] Step 432, the candidate intent categories corresponding to at least one selected candidate intent label are integrated according to preset integration rules, and the target intent category generated by the integration is used as the intent category corresponding to the target intent label, wherein the current scheduling intention information includes the target intent category, and the integration rules include one of the following: intent merging, intent splitting.

[0071] In this embodiment, a context information may contain multiple intents, or an intent may be split into multiple sub-intents. The large model can merge or split the intents according to predefined rules or model outputs to answer the target object's questions in more detail. That is, after selecting at least one candidate intent label, the selected candidate intent categories will be merged or split to generate the final recognition result, that is, the current scheduling intention information including the target intent category will be obtained.

[0072] Through the above steps 41 to 43, it is possible to determine the current scheduling intention information of the target object from the context information of the target object's interaction through natural language processing and intention recognition, that is, to determine the current scheduling intention, thereby providing a reference for matching the corresponding mechanism knowledge to generate a target APS that meets the expectations of the target object.

[0073] In some embodiments, current planning and scheduling intention information, multi-source production data, and target mechanism knowledge are input into the APS building module corresponding to the intelligent agent, and the target APS is generated using the retrieval-enhanced generation method to obtain advanced planning and scheduling results, including the following steps:

[0074] Step 51 : Using a preset context encoder, the received current scheduling intention information and the target mechanism knowledge are mapped and encoded one by one to generate a corresponding second context.

[0075] In this embodiment, the intention in the current scheduling intention information (including the expected target items, constraints and expected selected intelligent optimization algorithms for the current scheduling) and the knowledge items corresponding to the target mechanism knowledge are contextually combined. For example, if the current intention includes the intention to minimize order delays, the corresponding second context is: A0, natural language "minimize order delays", Q: minimize order delay target item, A1, formula " ”, which means “Minimize order delays when scheduling production from high to low levels and some orders are delayed.”

[0076] Step 52 : Based on the multi-source production data, all parameters in the target mechanism knowledge corresponding to the second context are converted into real numbers to generate a parameterized context.

[0077] In this embodiment, multi-source production data corresponding to the multi-resource constrained manufacturing system is synchronously acquired in a data call manner, and parameters of the second context are assigned, that is, the second context is converted into real numbers.

[0078] In step 53, the parameterized context is used as a prompt word, and the APS building module is used to perform retrieval enhancement generation processing to output the target APS. The APS building module is an advanced planning and scheduling generation model trained by a deep learning neural network based on an attention mechanism and knowledge fusion, and is trained to output advanced planning and scheduling corresponding to the sample context based on the input sample context.

[0079] In this embodiment, under the influence of prompt words composed of instructions, second context, and multi-source production data as input data, the big model conducts in-depth analysis and understanding of the parameterized context; when learning the parameterized context, the big model identifies key objectives, constraints, and related information of the intelligent optimization algorithm, and performs further semantic analysis and logical reasoning on the relevant corresponding information. For example, the big model may find conflicts or associations between target items and constraint items in multiple knowledge items, thereby providing a basis for generating more targeted answers; in this embodiment, after organizing and generating the corresponding parameterized context, a pre-trained neural network is used as an APS building module to perform retrieval augmented generation (RAG) to prompt the big model to learn the knowledge of specific scenario decision mechanisms, thereby generating the corresponding target APS.

[0080] In this embodiment, the APS building block is an advanced planning and scheduling generation model based on deep learning neural network training using attention mechanisms and knowledge fusion. During the training process, forward propagation, attention mechanism training, and context management are performed, specifically including:

[0081] 1. Forward Propagation

[0082] The input content is converted into a tensor format acceptable to the model and input into the hidden layer of the model. Complex nonlinear transformations are performed through neurons in multiple hidden layers. Each neuron receives the output from the previous layer, calculates the weighted sum and applies an activation function (such as ReLU, Sigmoid, etc.). In this process, the model extracts the features of the input data and gradually builds a more advanced semantic representation.

[0083] 2. Attention Mechanism

[0084] The model generates an attention score by calculating the dot product or similarity between the query vector and the key vector. Based on the attention score, weights are assigned to different input parts, allowing the model to focus more on important information. In this embodiment, the corresponding model better understands the semantics of the input content and generates more accurate output.

[0085] 3. Context Management

[0086] The model maintains a context window that stores context information related to the current task.

[0087] 4. Generate output

[0088] Based on the hidden layer's output and contextual information, the output layer generates the final output—the complete mathematical model, including the objective function and constraints. The generated text or structured output is decoded and restructured to make it more suitable for human reading habits and understanding. For example, the generated mathematical model formulas are arranged and organized according to a specific format.

[0089] In some embodiments, after the target APS is generated, the following steps are further performed: decoding and reassembling the target APS, and converting the generated target data through a LaTeX editor to generate a LaTeX file.

[0090] In some embodiments, building a mechanism knowledge base includes the following steps:

[0091] Step 71, determine the scheduling scenario characteristic parameters of the multi-resource constraint manufacturing system under the preset scheduling scenario, and write the initial sub-element parameter items corresponding to each preset scheduling scenario based on the scheduling scenario characteristic parameters, wherein the APS sub-objective function corresponding to the initial sub-element parameter item includes at least one of the following: minimizing the maximum completion time, minimizing order delays, minimizing the cleaning time and preset time after equipment switching, and maximizing capacity utilization; the APS sub-constraint parameters corresponding to the initial sub-element parameter item include at least one of the following: order level sorting constraint, delivery day sorting constraint, product inventory constraint, equipment production line capacity range and order integration or splitting constraint, raw material inventory constraint, equipment occupancy constraint, and each initial sub-element parameter item is associated with a knowledge text representing the parameters and variables of the corresponding parameter item.

[0092] In this embodiment, the corresponding mechanism knowledge base is constructed by first compiling parameter items with target items, constraint items and variation items as sub-elements according to the different scenarios and computing power conditions of the multi-resource constrained manufacturing system, and building an initial APS model library; in this embodiment, some parameters and variables in the initial APS model library include the following: total number of orders I; order number i, i=1,2,…,I; total number of employees H; employee number h, h=1,2,…,H; total number of equipment W, equipment number w, w=1,2,…,W; total number of product types V, product type number v, v=1,2,…,V; level R of order i i , R i is a real value. The larger the real value, the higher the R i The higher the arrival time of order i ; Delivery time for order i ; The equipment preparation time or fault repair time F required to start production of product v of order i wiv ; Quantity of product v required for order i CO iv ; The quantity of product v that needs to be scheduled for order i after considering inventory ; Order i is the quantity of product v required to be scheduled for the virtual order corresponding to the minimum production demand of the reactor ; The actual number of product v that needs to be scheduled for order i, n iv ; Indicates the nth product v that needs to be scheduled for order i; the raw materials M required for a single piece of product v v ; Inventory quantity M of raw materials for product v bv ; Raw material inventory is insufficient to meet the actual production demand of order i product v, so cycle F is used miv ; Inventory quantity B of product v v ; [B vmin ,B vmax ] represents the inventory range of product v; represents the capacity range of equipment w producing product v; the total number of production lines L; the production line number l, l = 1, 2, ..., L; represents the production capacity range of product v produced by production line l; the production time T of a single product v produced by production line l lv ; The production time T of equipment w producing a single product v wv ; Cleaning time F required when equipment switches from production v to production w ; The completion time of the last task of order i C i ; Indicates product The time when production starts and ends on production line 1; Indicates product The time when production starts and ends at equipment w; Indicates product Preset the start and end time before the equipment w is put into production; l (w) indicates whether production line l can be produced by equipment w. If yes, x l (w)=1, otherwise, x l (w)=0; Indicates whether the device w needs to be preset. If necessary, ,otherwise, ; Indicates whether the device w needs to be cleaned. If yes, ,otherwise, ; g(w,l) indicates whether the device w is working in front of the production line l. If the device w is working in front of the production line l, g(w,l)=1, otherwise, g(w,l)=0; a is a large positive number; Q0 and Q1 represent the product Production weighting factor taking into account the type of staff scheduling (night shift / non-night shift).

[0093] In this embodiment, the initial model library uses target items, constraints, and variant items as sub-elements. It sets minimizing the maximum completion time, minimizing order delays, minimizing the cleaning time and preset time after equipment switching, and maximizing capacity utilization as target items. It sets order level sorting constraints, delivery day sorting constraints, product inventory constraints, equipment production line capacity range and order integration or splitting constraints, raw material inventory constraints, and equipment occupancy constraints as constraints. Some of the target items and constraints of the initial model library are shown in Tables 1 and 2 below:

[0094] Table 1

[0095]

[0096] Table 2

[0097]

[0098] In this embodiment, the above parameters, variables, target items, constraint items and variant items are used to form an advanced planning and scheduling initial model library.

[0099] In step 72, after using the variable replacement method to replace the preset variables in the initial sub-element parameter items and generate new sub-element parameter items, the new sub-element parameter items and the initial sub-element parameter items are merged to generate an initial APS knowledge base, wherein the initial APS knowledge base includes candidate sub-element parameter items.

[0100] In this embodiment, the variable replacement method is used to enrich the constraint items. The corresponding construction enrichment example is as follows: define the equipment w production Required time limit, variant 1. ; Variant 2. .

[0101] Step 73, based on all candidate sub-element parameter items, the large model is stimulated by preset prompt words to perform knowledge expansion processing, and alternative sub-element parameter items are generated. All alternative sub-element parameter items are disassembled in a preset question-and-answer format to generate knowledge item pairs consisting of index text and knowledge text, wherein the knowledge expansion processing modifies one of the following target parameters: name, complexity level, mathematical formula, formula description, target parameters include objective function, constraint parameters, constraint variant item parameters, index text is used to represent the target corresponding to an alternative sub-element parameter item, and knowledge text is used to represent the parameters and variables corresponding to an alternative sub-element parameter item.

[0102] In this embodiment, the preset prompt words are used to stimulate the large model to assist in expanding the APS model library. The corresponding expansion methods include: modifying the names of targets and constraints, simplifying or complicating mathematical formulas, simulating the expression habits of different users to change the remark content, and ensuring that the core logic is consistent with the original intention; specifically including: 1. Writing prompt words; 2. Based on the original APS modeling target items, constraint items, and constraint variant items, the large model parameters receive prompt word inputs, and use the neural network layer to embed the prompt words and convert them into high-dimensional vectors. Then, information matching and transformation are performed through the multi-layer attention mechanism to generate extended content that meets the requirements; 3. The extended content includes modifying the names of target items, constraint items, and constraint variant items, simplifying or complicating the mathematical formulas of target items, constraint items, and constraint variant items, and simulating the expression habits of different users to change the corresponding explanatory content of the formula; 4. The extended content is appended to the end of the file. In this way, the content of the modeling knowledge file is expanded to make it a complete APS model library; in this embodiment, the mathematical formulas contained in the target items, constraint items and constraint variant items in the complete APS model library are all parsed into original character sequences through a format conversion tool, the grammatical structure of the formula is processed, and an appropriate LaTeX command sequence is generated according to the hierarchical relationship of the formula elements. It is reorganized into a LaTeX format formula according to the LaTeX format rules, and the converted LaTeX format formula replaces the original formula and is stored in the file. The computer overwrites the formula area in the file according to the character replacement rules to ensure that the converted format is compatible with the file context content format; in this embodiment, the mechanism knowledge of the complete APS model library is disassembled into the modeling knowledge file in the form of "QA (question-answer)", and after manual or semi-automatic reconstruction of the knowledge, the remaining knowledge reconstruction work is completed based on the large model to obtain the modeling mechanism knowledge.

[0103] Step 74, vectorize the knowledge item pairs corresponding to all the alternative sub-element parameter items, generate knowledge items corresponding to each alternative sub-element parameter item, and construct all the generated knowledge items into a mechanism knowledge base, wherein the vectorization processing includes supervised block processing, data index construction processing, feature extraction and alignment processing, and structure standardization processing.

[0104] In this embodiment, the mechanism knowledge in the constructed complete APS model library will be stored in the vector database after supervised segmentation, data index construction, feature extraction alignment, structural standardization and other processes. Among them, during supervised segmentation, the knowledge will be segmented according to custom rules (that is, according to the delimiters pre-set in the modeling knowledge file); then, the document is feature extracted by a machine learning algorithm to extract the keyword features of the knowledge block. At the same time, the feature vectors of different knowledge blocks are aligned to the same dimension, for example, by padding, cutting or dimensionality reduction methods to ensure that all feature vectors have the same length for subsequent vector operations. At the same time, the feature vectors are normalized so that they are distributed within the same numerical range; in the process of building the data index, a unique identifier is generated for each knowledge block and stored in association with the corresponding feature vector; the format of the knowledge data is adjusted to meet the storage requirements of the vector database, for example, the metadata (such as creation time, update time, etc.) and feature vector of each knowledge data block are stored as records in the database in a fixed format.

[0105] In this embodiment, based on the complete model library, guided by the ideas of intelligent optimization algorithms such as simulated annealing algorithm, genetic algorithm, particle swarm algorithm, etc., model instance solving is carried out, and an algorithm library is created in the form of Python programming code as output; in this embodiment, the mathematical formulas contained in the corresponding target items, constraint items and variant items in the model library are presented in LaTeX format with the help of a format conversion tool; in this embodiment, after constructing a complete APS model library, the target items and constraint items in the APS model library are disassembled in the form of questions and answers, and then the knowledge is reconstructed to obtain mechanism knowledge; the corresponding mechanism knowledge is stored in the corresponding database after supervised segmentation, data index construction, feature extraction alignment, structural standardization and other processing.

[0106] This embodiment also provides an advanced planning and scheduling generation device assisted by a large model intelligent agent, which is used to implement the above-mentioned embodiments and preferred implementation methods, and will not be repeated here. As used below, the terms "module", "unit", "sub-unit", etc. can implement a combination of software and / or hardware that implements predetermined functions. Although the devices described in the following embodiments are preferably implemented in software, implementation in hardware, or a combination of software and hardware, is also possible and conceivable.

[0107] Figure 3 This is a structural block diagram of an advanced planning and scheduling device based on a large model agent according to an embodiment of the present application, such as Figure 3 As shown, the device includes an acquisition module 31, a retrieval module 32 and a processing module 33, wherein:

[0108] The acquisition module 31 is used to obtain the current scheduling intention information expected by the target object from the context information generated by the natural language interaction with the target object, wherein the current scheduling intention information includes the sub-element parameter items associated with the current APS, and the sub-element parameter items include at least one of the following: APS sub-objective function, APS sub-constraint parameter, and an APS represents an advanced plan and scheduling.

[0109] The retrieval module 32 is coupled to the acquisition module 31 and is used to retrieve the target mechanism knowledge corresponding to the current scheduling intention information from the mechanism knowledge stored in the mechanism knowledge base corresponding to the preset intelligent agent based on the current scheduling intention information, wherein the mechanism knowledge base is generated based on the preset sub-element parameter items and the prompt words are used to stimulate the large model for knowledge expansion, and the preset sub-element parameter items are written and generated based on the scheduling scenario characteristic parameters of the multi-resource constraint manufacturing system, and each mechanism knowledge is associated with a sub-element parameter item.

[0110] The processing module 33 is coupled to the acquisition module 31 and the retrieval module 32. After acquiring the multi-source production data currently corresponding to the multi-resource constrained manufacturing system, it inputs the current planning and scheduling intention information, multi-source production data and target mechanism knowledge into the APS construction module corresponding to the intelligent agent, generates the target APS using the retrieval enhanced generation method, and obtains advanced planning and scheduling results, wherein the multi-source production data is used to characterize the multi-dimensional production status corresponding to the multi-resource constrained manufacturing system.

[0111] In some embodiments, the retrieval module 32 further includes:

[0112] The first generation unit is used to perform natural language processing and word vector processing on the current scheduling intention information, generate an intention vector corresponding to the current scheduling intention information, and determine the knowledge entry corresponding to each mechanism knowledge, wherein the knowledge entry includes a first context composed of an index vector and a knowledge text vector, one index vector corresponds to a sub-element parameter item, and the knowledge text vector is used to represent the parameters and variables of the associated index vector.

[0113] The computing unit is coupled to the first generating unit and is used to determine the similarity between the index vectors corresponding to all knowledge items and the intention vector, and select a preset number of index vectors in descending order of similarity to obtain a target index vector.

[0114] The processing unit is coupled to the computing unit and is used to use the knowledge entry corresponding to the target index vector as the target mechanism knowledge corresponding to the current scheduling intention information.

[0115] In some embodiments, the generation unit is also used to perform natural language processing on the current scheduling intention information to generate intention vocabulary corresponding to the current scheduling intention information; perform word embedding processing and word mapping processing on the intention vocabulary in turn to generate corresponding word vectors, wherein the word vector is at least used to represent the semantic information of the corresponding intention vocabulary; use a preset context encoder to combine all word vectors into a context semantic vector, wherein the intention vector includes a context semantic vector.

[0116] In some embodiments, the acquisition module 31 further includes:

[0117] The encoding unit is used to process the context information generated by the target object during the interaction process using natural language processing (NLP) to obtain encoded data. The context information includes the target parameters of the current scheduling, constraint requirements and the intelligent optimization algorithm to be selected. The intelligent optimization algorithm includes one of the following: simulated annealing algorithm, genetic algorithm, particle swarm algorithm. The data processing includes word segmentation, stop word removal and vocabulary conversion encoding.

[0118] The recognition unit is coupled to the encoding unit and is used to use a preset intent recognition network to perform feature extraction and intent recognition on the encoded data to generate multiple candidate intent labels, where the candidate intent labels include candidate intent categories and candidate intent category confidences. The intent recognition network is an intent recognition model trained using a deep learning neural network based on an attention mechanism, and is trained to identify intent labels corresponding to the sample question and answer information based on the input sample question and answer information.

[0119] The selection unit is coupled to the recognition unit and is used to select a target intention label from multiple candidate intention labels based on the confidence of the candidate intention category to obtain the current scheduling intention information.

[0120] In some embodiments, the selection unit is also used to select at least one candidate intent label whose candidate intent category confidence is greater than a confidence threshold from multiple candidate intent labels to obtain a target intent label; the candidate intent category corresponding to the selected at least one candidate intent label is integrated according to a preset integration rule, and the target intent category generated by the integration is used as the intent category corresponding to the target intent label, wherein the current scheduling intention information includes the target intent category, and the integration rule includes one of the following: intent merging and intent splitting.

[0121] In some embodiments, the processing module 33 further includes:

[0122] The second generating unit is used to use a preset context encoder to perform one-to-one mapping encoding on the received current scheduling intention information and the target mechanism knowledge to generate a corresponding second context.

[0123] The third generating unit is coupled to the second generating unit and is used to perform real number processing on parameters in the target mechanism knowledge corresponding to all the second contexts based on multi-source production data to generate a parameterized context.

[0124] The solving unit is coupled to the third generating unit and is used to use the parameterized context as a prompt word, utilize the APS building module to perform retrieval enhancement generation processing, and output the target APS, wherein the APS building module is an advanced planning and scheduling generation model trained by a deep learning neural network based on the attention mechanism and knowledge fusion, and is trained to output advanced planning and scheduling corresponding to the sample context based on the input sample context.

[0125] In some embodiments, after generating the target APS, the processing module 33 is further configured to decode and reconstruct the target APS, and convert the generated target data through a LaTeX editor to generate a LaTeX file.

[0126] In some embodiments, the advanced planning and scheduling device assisted by a large model agent further constructs a mechanism knowledge base by:

[0127] Determine the scheduling scenario characteristic parameters of the multi-resource constrained manufacturing system under the preset scheduling scenario, and compile the initial sub-element parameter items corresponding to each preset scheduling scenario based on the scheduling scenario characteristic parameters, wherein the APS sub-objective function corresponding to the initial sub-element parameter item includes at least one of the following: minimizing the maximum completion time, minimizing order delays, minimizing the cleaning time and preset time after equipment switching, and maximizing capacity utilization; the APS sub-constraint parameters corresponding to the initial sub-element parameter item include at least one of the following: order level sorting constraint, delivery day sorting constraint, product inventory constraint, equipment production line capacity range and order integration or splitting constraint, raw material inventory constraint, equipment occupancy constraint, and each initial sub-element parameter item is associated with a knowledge text representing the parameters and variables of the corresponding parameter item;

[0128] After replacing the preset variables in the initial sub-element parameter items by using the variable replacement method to generate new sub-element parameter items, the new sub-element parameter items and the initial sub-element parameter items are merged to generate an initial APS knowledge base, wherein the initial APS knowledge base includes candidate sub-element parameter items;

[0129] Based on all candidate sub-element parameter items, a large model is stimulated with preset prompt words to perform knowledge expansion processing to generate alternative sub-element parameter items. All alternative sub-element parameter items are disassembled according to a preset question-and-answer format to generate knowledge item pairs consisting of index text and knowledge text. The knowledge expansion processing modifies one of the following target parameters: name, complexity level, mathematical formula, and formula description. The target parameters include objective function, constraint parameters, and constraint variant item parameters. The index text is used to represent the target corresponding to an alternative sub-element parameter item, and the knowledge text is used to represent the parameters and variables corresponding to an alternative sub-element parameter item.

[0130] The knowledge entry pairs corresponding to all the alternative sub-element parameter items are vectorized to generate knowledge entries corresponding to each alternative sub-element parameter item, and all the generated knowledge entries are constructed into a mechanism knowledge base, wherein the vectorization processing includes supervised block processing, data index construction processing, feature extraction and alignment processing, and structure standardization processing.

[0131] This embodiment also provides an advanced planning and scheduling system, including a memory and a processor, wherein the memory stores a computer program, and the processor is configured to run the computer program to execute the steps in any of the above method embodiments.

[0132] Optionally, the service platform may further include a transmission device and an input / output device, wherein the transmission device is connected to the processor, and the input / output device is connected to the processor.

[0133] Optionally, in this embodiment, the processor may be configured to execute the following steps through a computer program:

[0134] S1. Obtain the current scheduling intention information expected by the target object from the context information generated by natural language interaction with the target object. The current scheduling intention information includes the sub-element parameter items associated with the current APS. The sub-element parameter items include at least one of the following: APS sub-objective function, APS sub-constraint parameter. An APS represents an advanced plan and scheduling.

[0135] S2, based on the current scheduling intention information, retrieves the target mechanism knowledge corresponding to the current scheduling intention information from the mechanism knowledge stored in the mechanism knowledge base corresponding to the preset intelligent agent. The mechanism knowledge base is generated based on the preset sub-element parameter items and the prompt words are used to stimulate the large model for knowledge expansion. The preset sub-element parameter items are written and generated based on the scheduling scenario characteristic parameters of the multi-resource constraint manufacturing system. Each mechanism knowledge is associated with a sub-element parameter item.

[0136] S3, after obtaining the multi-source production data currently corresponding to the multi-resource constrained manufacturing system, input the current planning and scheduling intention information, multi-source production data and target mechanism knowledge into the APS building module corresponding to the intelligent agent, and use the retrieval enhancement generation method to generate the target APS to obtain advanced planning and scheduling results. The multi-source production data is used to characterize the multi-dimensional production status corresponding to the multi-resource constrained manufacturing system.

[0137] It should be noted that the specific examples in this embodiment can refer to the examples described in the above embodiments and optional implementation modes, and this embodiment will not be repeated here.

[0138] In addition, in conjunction with the large-model agent-assisted advanced planning and scheduling generation method in the above-mentioned embodiments, embodiments of the present application may provide a storage medium for implementation. The storage medium stores a computer program; when executed by a processor, the computer program implements any of the large-model agent-assisted advanced planning and scheduling generation methods in the above-mentioned embodiments.

[0139] Those skilled in the art should understand that the various technical features of the above embodiments can be combined arbitrarily. To make the description concise, not all possible combinations of the various technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0140] The above embodiments merely represent several implementation methods of the present application. While the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art could make various modifications and improvements without departing from the spirit of the present application, all of which fall within the scope of protection of the present application. Therefore, the scope of protection of the present patent application shall be determined by the appended claims.

Claims

1. A large-scale intelligent agent-assisted advanced planning and scheduling generation method, characterized in that: include: Obtaining the target object's expected current scheduling intention information from the context information generated by natural language interaction with the target object, wherein the current scheduling intention information includes sub-element parameter items associated with the current APS, and the sub-element parameter items include at least one of the following: APS sub-objective function, APS sub-constraint parameter, and an APS represents a type of advanced planning and scheduling; Based on the current scheduling intention information, target mechanism knowledge corresponding to the current scheduling intention information is retrieved from the mechanism knowledge stored in the mechanism knowledge base corresponding to the preset intelligent agent, wherein the mechanism knowledge base is generated by using prompt words to stimulate the large model for knowledge expansion based on the preset sub-element parameter items, and the preset sub-element parameter items are compiled and generated based on the scheduling scenario characteristic parameters of the multi-resource constrained manufacturing system, and each mechanism knowledge is associated with one sub-element parameter item; After obtaining the current multi-source production data corresponding to the multi-resource constrained manufacturing system, the current planning and scheduling intention information, the multi-source production data, and the target mechanism knowledge are input into the APS building module corresponding to the intelligent agent, and the target APS is generated using the retrieval-enhanced generation method to obtain an advanced planning and scheduling result, wherein the multi-source production data is used to represent the multi-dimensional production status corresponding to the multi-resource constrained manufacturing system; The current planning and scheduling intention information, the multi-source production data and the target mechanism knowledge are input into the APS building module corresponding to the agent, and the target APS is generated using the retrieval enhancement generation method to obtain advanced planning and scheduling results, including: Using a preset context encoder, the received current scheduling intention information and the target mechanism knowledge are mapped and encoded one by one to generate a corresponding second context; Based on the multi-source production data, all parameters in the target mechanism knowledge corresponding to the second context are converted into real numbers to generate a parameterized context; Using the parameterized context as a prompt word, the APS building module is used to perform retrieval enhancement generation processing and output the target APS, wherein the APS building module is an advanced planning and scheduling generation model trained by a deep learning neural network based on an attention mechanism and knowledge fusion, and is trained to output an advanced plan and scheduling corresponding to the sample context based on the input sample context; The construction of the mechanism knowledge base includes: Determine the scheduling scenario characteristic parameters of the multi-resource constraint manufacturing system under a preset scheduling scenario, and compile initial sub-element parameter items corresponding to each preset scheduling scenario based on the scheduling scenario characteristic parameters, wherein the APS sub-objective function corresponding to the initial sub-element parameter item includes at least one of the following: minimizing maximum completion time, minimizing order delays, minimizing cleaning time and preset time after equipment switching, and maximizing capacity utilization; the APS sub-constraint parameters corresponding to the initial sub-element parameter item include at least one of the following: order level ranking constraint, delivery day ranking constraint, product inventory constraint, equipment production line capacity range and order integration or splitting constraint, raw material inventory constraint, and equipment occupancy constraint; each initial sub-element parameter item is associated with a knowledge text representing the parameters and variables of the corresponding parameter item; After replacing the preset variables in the initial sub-element parameter items using a variable replacement method to generate new sub-element parameter items, the new sub-element parameter items are merged with the initial sub-element parameter items to generate an initial APS knowledge base, wherein the initial APS knowledge base includes candidate sub-element parameter items; Based on all the candidate sub-element parameter items, a large model is stimulated with preset prompt words to perform knowledge expansion processing to generate candidate sub-element parameter items, and all the candidate sub-element parameter items are disassembled in a preset question-and-answer format to generate knowledge item pairs consisting of index text and knowledge text, wherein the knowledge expansion processing modifies one of the following target parameters: name, complexity level, mathematical formula, and formula description; the target parameters include objective functions, constraint parameters, and constraint variant item parameters; the index text is used to represent the target corresponding to one of the candidate sub-element parameter items, and the knowledge text is used to represent the parameters and variables corresponding to one of the candidate sub-element parameter items; The knowledge item pairs corresponding to all the alternative sub-element parameter items are vectorized to generate the knowledge item corresponding to each of the alternative sub-element parameter items, and all the generated knowledge items are constructed into the mechanism knowledge base, wherein the vectorization processing includes supervised block processing, data index construction processing, feature extraction and alignment processing, and structural standardization processing.

2. The method according to claim 1, characterized in that Based on the current scheduling intention information, searching for target mechanism knowledge corresponding to the current scheduling intention information from the mechanism knowledge stored in the mechanism knowledge base corresponding to the preset intelligent agent includes: Performing natural language processing and word vector processing on the current scheduling intention information to generate an intention vector corresponding to the current scheduling intention information, and determining a knowledge item corresponding to each of the mechanism knowledge items, wherein the knowledge item includes a first context consisting of an index vector and a knowledge text vector, one index vector corresponds to one sub-element parameter item, and the knowledge text vector is used to represent parameters and variables of the associated index vector; Determine the similarity between the index vectors corresponding to all the knowledge items and the intention vector, and select a preset number of the index vectors in descending order of similarity to obtain a target index vector; The knowledge entry corresponding to the target index vector is used as the target mechanism knowledge corresponding to the current scheduling intention information.

3. The method according to claim 2, characterized in that Performing natural language processing and word vector processing on the current scheduling intention information to generate an intention vector corresponding to the current scheduling intention information includes: Performing natural language processing on the current metering intention information to generate an intention vocabulary corresponding to the current metering intention information; Performing word embedding and word mapping processing on the intended vocabulary in sequence to generate corresponding word vectors, wherein the word vectors are at least used to represent the semantic information of the corresponding intended vocabulary; Using a preset context encoder, all of the word vectors are combined into a context semantic vector, wherein the intention vector includes the context semantic vector.

4. The method according to claim 2, characterized in that Obtaining the target object's expected current scheduling intention information from the context information generated by the natural language interaction with the target object, including: The context information generated by the target object during the interaction process is processed using natural language processing (NLP) to obtain encoded data, wherein the context information includes target parameters, constraint requirements, and a proposed intelligent optimization algorithm for current scheduling, wherein the intelligent optimization algorithm includes one of the following: simulated annealing algorithm, genetic algorithm, and particle swarm algorithm; and the data processing includes word segmentation, stop word removal, and vocabulary conversion encoding; Using a preset intent recognition network, feature extraction and intent recognition are performed on the encoded data to generate multiple candidate intent labels, where the candidate intent labels include candidate intent categories and candidate intent category confidences. The intent recognition network is an intent recognition model trained using a deep learning neural network based on an attention mechanism and is trained to identify intent labels corresponding to sample question and answer information based on the input sample question and answer information; According to the confidence of the candidate intention category, a target intention label is selected from the multiple candidate intention labels to obtain the current scheduling intention information.

5. The method according to claim 4, characterized in that According to the candidate intention category confidence, a target intention label is selected from the plurality of candidate intention labels to obtain the current scheduling intention information, including: Selecting at least one candidate intent label whose candidate intent category confidence is greater than a confidence threshold from among the plurality of candidate intent labels to obtain the target intent label; The candidate intent categories corresponding to at least one selected candidate intent label are integrated according to preset integration rules, and the target intent category generated by the integration is used as the intent category corresponding to the target intent label, wherein the current scheduling intention information includes the target intent category, and the integration rules include one of the following: intent merging and intent splitting.

6. The method according to claim 1, characterized in that After generating the target APS, the method further includes: decoding and reorganizing the target APS, and converting the generated target data through a LaTeX editor to generate a LaTeX file.

7. An advanced planning and scheduling system, comprising a memory and a processor, characterized in that: A computer program is stored in the memory, and the processor is configured to run the computer program to perform the steps of the large model agent-assisted advanced planning and scheduling generation method according to any one of claims 1 to 6.

8. A storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the large model agent-assisted advanced planning and scheduling generation method according to any one of claims 1 to 6 are implemented.

Images