Complex manufacturing system production scheduling engine based on large language model spatio-temporal reasoning

Abstract

The invention provides a complex manufacturing system production scheduling engine based on large language model spatio-temporal reasoning, and relates to the field of intelligent manufacturing, a scheduling demand is obtained, time-related information and space-related information are extracted, then spatial features and a time constraint graph are jointly coded into LLM analyzable spatio-temporal semantic vectors, and the LLM analyzable spatio-temporal semantic vectors are used for scheduling. And explicitly modeling a multi-dimensional interaction relationship of spatial proximity-time overlapping degree-resource competition intensity through association analysis of scheduling space-time semantics, so that LLM can realize more reliable space-time causal reasoning, and finally a scheduling scheme conforming to complex space-time constraints is generated. Thus, high-dimensional spatial-temporal characteristics in a manufacturing system can be adapted, spatial-temporal information required by user scheduling can be highlighted to adapt to an input format of LLM, then a nonlinear spatial-temporal coupling relationship in the production scheduling of a complex manufacturing system is concerned, and the generalization ability is improved.

Description

Technical Field

[0001] This application relates to the field of intelligent manufacturing, and more particularly, in the embodiments of this application, it relates to a production scheduling engine for complex manufacturing systems based on spatio-temporal reasoning of large language models. Background Art

[0002] The production scheduling of complex manufacturing systems is a core issue in the field of intelligent manufacturing. Its goal is to maximize resource utilization, minimize lead time, and optimize production costs in a dynamic environment with multiple processes, multiple devices, and multiple constraints. Traditional production scheduling methods mainly rely on mathematical programming (such as mixed integer programming), heuristic algorithms (such as genetic algorithms, simulated annealing), or rule-based systems, but they face significant challenges in dealing with complex spatio-temporal coupling relationships: the spatial layout of the manufacturing system (such as the physical location of equipment, material handling paths) and time constraints (such as process sequence, time windows) are strongly coupled. Traditional methods usually decouple them into independent optimization objectives, resulting in suboptimal solutions or constraint conflicts. In addition, existing models are difficult to dynamically represent the non-linear relationships among equipment idle time, material waiting queues, and spatial movement paths. For example, long-distance material handling may offset the time gains of parallel processes. Moreover, numerically optimized methods cannot directly analyze implicit spatio-temporal semantic logics such as "adjacent areas lead to shorter mold change times" and "path intersections cause AGV congestion risks", relying on the construction of artificial rule bases and having limited generalization capabilities.

[0003] Therefore, an optimized production scheduling engine for complex manufacturing systems is desired. Summary of the Invention

[0004] To solve the above technical problems, this application is proposed. The embodiments of this application provide a production scheduling engine for complex manufacturing systems based on spatio-temporal reasoning of large language models. It first obtains scheduling requirements and extracts time-related information and space-related information. Then, it jointly encodes spatial features and time constraint graphs into spatio-temporal semantic vectors that can be parsed by the LLM, and explicitly models the multi-dimensional interaction relationship of "spatial proximity - time overlap - resource competition intensity" through the correlation analysis of scheduling spatio-temporal semantics, enabling the LLM to achieve more reliable spatio-temporal causal reasoning. Finally, it generates a scheduling plan that meets complex spatio-temporal constraints. In this way, it can adapt to the high-dimensional spatio-temporal features in the manufacturing system, highlight the spatio-temporal information of user scheduling requirements to adapt to the input format of the LLM, and then focus on the non-linear spatio-temporal coupling relationships in the production scheduling of complex manufacturing systems, improving the generalization ability.

[0005] According to one aspect of this application, there is provided a production scheduling engine for complex manufacturing systems based on spatio-temporal reasoning of large language models, which includes:

[0006] The scheduling requirement decomposition module is used to obtain the scheduling requirements input by the user, and extract the scheduling basic information part, the scheduling requirement time-related part, and the scheduling requirement space-related part from the scheduling requirements;

[0007] The scheduling requirement spatio-temporal semantic interaction correlation encoding module is used to perform scheduling spatio-temporal semantic correlation analysis on the scheduling requirement time-related part and the scheduling requirement space-related part based on the semantic feature space to obtain the scheduling requirement spatio-temporal part interaction correlation representation information;

[0008] The scheduling plan generation module is used to determine the scheduling plan based on the scheduling requirement spatio-temporal part interaction correlation representation information and the scheduling basic information part, and use the large language model;

[0009] Among them, the scheduling requirement spatio-temporal semantic interaction correlation encoding module is used to: perform interactive parsing based on the scheduling spatio-temporal performance correlation on the scheduling requirement time-related part and the scheduling requirement space-related part to obtain the scheduling requirement spatio-temporal part interaction correlation representation information.

[0010] Compared with the prior art, a complex manufacturing system production scheduling engine based on large language model spatio-temporal reasoning provided by the present application first obtains scheduling requirements and extracts time-related information and space-related information. Then, it jointly encodes the space features and the time constraint graph into spatio-temporal semantic vectors that can be parsed by the LLM, and explicitly models the multi-dimensional interaction relationship of "spatial proximity - time overlap degree - resource competition intensity" through the correlation analysis of the scheduling spatio-temporal semantics, enabling the LLM to achieve more reliable spatio-temporal causal reasoning, and finally generating a scheduling plan that meets complex spatio-temporal constraints. In this way, it can adapt to the high-dimensional spatio-temporal features in the manufacturing system, highlight the spatio-temporal information of the user's scheduling requirements to adapt to the input format of the LLM, and then focus on the non-linear spatio-temporal coupling relationship in the complex manufacturing system production scheduling, improving the generalization ability. BRIEF DESCRIPTION OF THE DRAWINGS

[0011] By describing the embodiments of the present application in more detail in conjunction with the drawings, the above and other objects, features, and advantages of the present application will become more obvious. The drawings are used to provide a further understanding of the embodiments of the present application, and constitute a part of the specification. Together with the embodiments of the present application, they are used to explain the present application, and do not constitute a limitation to the present application. In the drawings, the same reference numerals generally represent the same components or steps.

[0012] Figure 1 It is a system block diagram of a complex manufacturing system production scheduling engine based on large language model spatio-temporal reasoning according to an embodiment of the present application.

[0013] Figure 2 It is a data flow diagram of a complex manufacturing system production scheduling engine based on large language model spatio-temporal reasoning according to an embodiment of the present application.

[0014] Figure 3 It is a block diagram of a scheduling requirement spatio-temporal semantic interaction and correlation encoding module in a production scheduling engine for a complex manufacturing system based on spatio-temporal reasoning of a large language model according to an embodiment of the present application.

[0015] Figure 4 It is a block diagram of a fine-grained interaction and correlation unit for the spatio-temporal part of scheduling requirements in a production scheduling engine for a complex manufacturing system based on spatio-temporal reasoning of a large language model according to an embodiment of the present application.

[0016] Figure 5 It is a block diagram of a scheduling plan generation module in a production scheduling engine for a complex manufacturing system based on spatio-temporal reasoning of a large language model according to an embodiment of the present application. Detailed implementation manners

[0017] Various exemplary embodiments, features, and aspects of the present application will be described in detail below with reference to the accompanying drawings. The same reference numerals in the drawings denote elements having the same or similar functions. Although various aspects of the embodiments are shown in the drawings, the drawings are not necessarily drawn to scale unless otherwise specified.

[0018] The special term "exemplary" here means "serving as an example, embodiment, or illustration". Any embodiment described as "exemplary" here need not be construed as superior to or better than other embodiments.

[0019] In addition, for a better illustration of the present application, numerous specific details are given in the following detailed implementation manners. Those skilled in the art should understand that the present application can be implemented without some specific details. In some instances, methods, means, elements, and circuits well known to those skilled in the art are not described in detail so as to highlight the gist of the present application.

[0020] Furthermore, the terms "first" and "second" are used for descriptive purposes only and cannot be construed as indicating or implying relative importance or implicitly specifying the quantity of the indicated technical features. Thus, features defined with "first" and "second" may explicitly or implicitly include one or more of such features. In the description of the present application, "a plurality" means two or more unless otherwise specifically defined.

[0021] The production scheduling of complex manufacturing systems aims to optimize resource utilization, delivery cycle, and production cost in a dynamic environment with multiple processes, multiple devices, and multiple constraints. However, traditional methods such as mathematical programming, heuristic algorithms, or rule-based systems face challenges in dealing with the strong coupling relationship between spatial layout and time constraints, often decoupling them into independent objectives, resulting in suboptimal solutions or constraint conflicts. In addition, existing models are difficult to dynamically reflect the non-linear relationship between equipment idle time, material waiting queues, and spatial movement paths, and cannot directly analyze the implicit spatio-temporal semantic logic brought about by regional adjacency or path crossing, limiting their generalization ability.

[0022] In recent years, large language models (LLMs) have demonstrated the ability to deeply associate unstructured semantic information in complex reasoning tasks. However, there are still some limitations in their application to production scheduling in the spatio-temporal scenarios of complex manufacturing systems: the native text input paradigm of LLMs is difficult to adapt to the high-dimensional spatio-temporal features in manufacturing systems (such as equipment orientation topology maps, time constraint directed graphs). It is difficult to highlight the spatio-temporal information in the scheduling requirements input by users, and this information is also difficult to adapt to the input format of LLMs, easily overlooking the spatio-temporal coupling relationship in the production scheduling of complex manufacturing systems (such as the chain offset of the downstream process time window caused by the space limitation of a certain workstation).

[0023] To address the above technical problems, in the technical solution of this application, a production scheduling engine for complex manufacturing systems based on spatio-temporal reasoning of large language models is proposed. It extracts time-related information and space-related information in the user's scheduling requirements and captures the spatio-temporal partial interaction information of the scheduling requirements, providing more effective data input for the large language model. Specifically, it can jointly encode spatial features such as equipment coordinates, distance and orientation relationships between equipment, and material flow topology with time constraint graphs into spatio-temporal semantic vectors that can be parsed by the LLM, and explicitly model the multi-dimensional interaction relationship of "spatial proximity - time overlap degree - resource competition intensity" through the correlation analysis of scheduling spatio-temporal semantics, enabling the LLM to achieve more reliable spatio-temporal causal reasoning, such as automatically identifying implicit conflicts such as "insufficient capacity of the in-process buffer caused by cross-regional equipment collaboration", and finally generating a scheduling plan that meets complex spatio-temporal constraints.

[0024] This application proposes a production scheduling engine for complex manufacturing systems based on spatio-temporal reasoning of large language models. Figure 1 It is a system block diagram of the production scheduling engine for complex manufacturing systems based on spatio-temporal reasoning of large language models according to an embodiment of this application. Figure 2 It is a schematic diagram of data flow of the production scheduling engine for complex manufacturing systems based on spatio-temporal reasoning of large language models according to an embodiment of this application. As Figure 1 and Figure 2As shown in the figure, the complex manufacturing system production scheduling engine 100 based on spatio-temporal reasoning of large language models according to an embodiment of the present application includes: a scheduling requirement decomposition module 110, configured to obtain scheduling requirements input by a user, and extract a scheduling basic information part, a scheduling requirement time-related part, and a scheduling requirement space-related part from the scheduling requirements; a scheduling requirement spatio-temporal semantic interaction correlation encoding module 120, configured to perform scheduling spatio-temporal semantic correlation analysis based on a semantic feature space on the scheduling requirement time-related part and the scheduling requirement space-related part to obtain scheduling requirement spatio-temporal part interaction correlation representation information; a scheduling plan generation module 130, configured to determine a scheduling plan based on the scheduling requirement spatio-temporal part interaction correlation representation information and the scheduling basic information part, and by using a large language model.

[0025] In the above-mentioned production scheduling engine 100 of the complex manufacturing system based on large language model spatio-temporal reasoning, the scheduling requirement decomposition module 110 is used to obtain the scheduling requirements input by the user, and extract the scheduling basic information part, the scheduling requirement time-related part, and the scheduling requirement space-related part from the scheduling requirements. It should be understood that obtaining scheduling requirements from the user captures and interprets the user's intentions. This involves clarifying specific production goals, such as basic information like order delivery time and product quality requirements, as well as time-related specific requirements, such as time windows for each process and equipment maintenance cycles. At the same time, considering the impact of factors such as equipment layout and material handling paths in the manufacturing system on scheduling, it is also necessary to accurately extract space-related information. Specifically, for the extraction of the scheduling basic information part, it is necessary to identify the basic elements that affect the formulation of the entire scheduling plan. These information usually include but are not limited to order details (such as product type, quantity), production process flow (which processes are involved and specific requirements for each process), resource availability (such as human resources, equipment status), etc. By comprehensively collecting these basic data, understanding the priority of each order and its deadline helps to determine the production sequence, ensuring that critical orders can be completed within the specified time, thus meeting customer needs and maintaining the market competitiveness of the enterprise. The scheduling requirement time-related part can capture all information closely related to the time dimension. This means not only considering the specific duration required for a single process or activity, but also paying attention to whether there are dependencies between different processes, that is, some processes must start after other processes are completed. Further, it is also necessary to evaluate the time fluctuations that may be caused by various external factors, such as unforeseen situations like delayed arrival of supplier materials and sudden equipment failures. Through in-depth analysis of these time-related factors, the production progress can be predicted more accurately, unnecessary waiting time can be reduced, and overall efficiency can be improved. The scheduling requirement space-related part focuses on the impact of the physical space layout on scheduling decisions. The equipment in the manufacturing system does not exist in isolation, and the relative position relationship between them and the material transportation path will directly affect the effectiveness and efficiency of scheduling. For example, in a large-scale production workshop, the material handling between two devices located far apart may consume a lot of time, thus affecting the speed of the entire production chain. Therefore, understanding factors such as the distance between devices, the design of the material flow line, and the storage location is crucial for formulating a reasonable scheduling plan. In addition, considering safety operation procedures and work environment limitations, there may be special access conditions or operation requirements in some areas, which also need to be considered during the scheduling process. In this way, obtaining the scheduling requirements input by the user and extracting the scheduling basic information part, the scheduling requirement time-related part, and the scheduling requirement space-related part from it can ensure that the finally generated scheduling plan not only meets the complex spatio-temporal constraints, but also maximally meets the user's needs, achieving the maximization of resource utilization rate, the minimization of delivery cycle, and the optimization of production cost.

[0026] In the above-mentioned complex manufacturing system production scheduling engine 100 based on large language model spatio-temporal reasoning, the scheduling requirement spatio-temporal semantic interaction and correlation encoding module 120 is used to perform scheduling spatio-temporal semantic correlation analysis based on the semantic feature space for the time-related part of the scheduling requirement and the space-related part of the scheduling requirement to obtain the scheduling requirement spatio-temporal part interaction and correlation representation information. Among them, the scheduling requirement spatio-temporal semantic interaction and correlation encoding module 120 is used to: perform interactive parsing based on the scheduling spatio-temporal performance correlation for the time-related part of the scheduling requirement and the space-related part of the scheduling requirement to obtain the scheduling requirement spatio-temporal part interaction and correlation representation information.

[0027] Figure 3 It is a block diagram of the scheduling requirement spatio-temporal semantic interaction and correlation encoding module in the complex manufacturing system production scheduling engine based on large language model spatio-temporal reasoning according to an embodiment of the present application. As Figure 3 shown, in the embodiment of the present application, the scheduling requirement spatio-temporal semantic interaction and correlation encoding module 120 includes: a scheduling requirement spatio-temporal semantic embedding and encoding unit 121, which is used to perform semantic embedding and encoding on the time-related part of the scheduling requirement and the space-related part of the scheduling requirement to obtain the semantic embedding and encoding features of the time-related part of the scheduling requirement and the semantic embedding and encoding features of the space-related part of the scheduling requirement; a scheduling requirement spatio-temporal part fine-grained interaction and correlation unit 122, which is used to perform fine-grained interaction analysis based on the scheduling spatio-temporal deep semantic feature performance correlation on the semantic embedding and encoding features of the time-related part of the scheduling requirement and the semantic embedding and encoding features of the space-related part of the scheduling requirement to obtain the scheduling requirement spatio-temporal part interaction and correlation representation information.

[0028] Specifically, the scheduling requirement spatio-temporal semantic embedding encoding unit 121 is used to perform semantic embedding encoding on the time-related part and the space-related part of the scheduling requirement to obtain the semantic embedding encoding features of the time-related part of the scheduling requirement and the semantic embedding encoding features of the space-related part of the scheduling requirement. It should be understood that considering the heterogeneity and coupling of spatio-temporal information in the production scheduling scenario of complex manufacturing systems, the traditional text input method is difficult to be directly understood by large language models (LLMs). Since the spatial features such as the equipment orientation and material path in the manufacturing system and the time features such as the process sequence and time window are essentially high-dimensional structured data, their topological relationships (such as the spatial graph formed by equipment coordinates and the directed graph formed by time constraints) cannot be completely transmitted through the description of the original text. For example, the implicit logics such as the quantitative relationship of the distance between equipment and the dynamic overlap of process sequences, if only input in the form of natural language, are easily simplified into discrete symbols by the general text processing mechanism of the LLM, losing the continuous correlation features of spatial proximity and time overlap. Therefore, in the technical solution of this application, the time-related part and the space-related part of the scheduling requirement are further subjected to semantic embedding encoding to obtain the semantic embedding encoding vectors of the time-related part of the scheduling requirement and the semantic embedding encoding vectors of the space-related part of the scheduling requirement. By converting the scheduling requirement spatio-temporal information into low-dimensional dense vectors through semantic embedding encoding, it is essentially a mathematical abstraction of the spatio-temporal coupling relationship of the manufacturing system. The spatial features of the scheduling requirement (such as regional adjacency relationship, handling path length) are extracted by the graph neural network to obtain topological embeddings, mapping the physical space of the equipment layout into the relative distance in the vector space; the time features of the scheduling requirement (such as the sequence before and after the process, the idle period of the equipment) are converted into semantic vectors with time dimension directivity through the time series encoder. This encoding process not only retains the mathematical representation ability of non-linear relationships such as "long-distance handling offsets the gain of parallel processes" and "the time window offset caused by the space limitation of the workbench", but also explicitly depicts the interaction intensity between the spatial orientation and time constraints through the geometric properties (such as cosine similarity) of the vector space. For example, when the spatial coordinate vector of a certain equipment and the time idle vector of another equipment show high correlation in the embedding space, it can reflect the potential contradiction of "spatially reachable but time-conflicting" between the two in the scheduling. In this way, not only the adaptation gap between the native text input of the LLM and the high-dimensional spatio-temporal data of the manufacturing system is solved, but also the basis for the subsequent spatio-temporal semantic fine-grained interaction correlation analysis of the scheduling requirement is provided.

[0029] Figure 4 FIG. is a block diagram of the spatio-temporal part fine-grained interaction correlation unit of the production scheduling engine of a complex manufacturing system based on spatio-temporal reasoning of a large language model according to an embodiment of the present application. As Figure 4As shown, in the embodiment of the present application, the fine-grained interaction correlation unit 122 of the scheduling requirement spatio-temporal part includes: a kernel correlation information encoding subunit 1221 of the scheduling requirement, which is used to perform kernel correlation information encoding based on principal component analysis on the semantic embedding encoding features of the time-related part of the scheduling requirement and the semantic embedding encoding vector features of the space-related part of the scheduling requirement to obtain a set of kernel correlation encoding vectors between the principal components of the spatio-temporal part features of the scheduling requirement; a correlation encoding subunit 1222 of the scheduling requirement, which is used to construct a kernel correlation performance analysis factor between every two kernel correlation encoding vectors between the principal components of the spatio-temporal part features of the scheduling requirement in the set of kernel correlation encoding vectors between the principal components of the spatio-temporal part features of the scheduling requirement, and perform graph structure-based correlation encoding on the set of kernel correlation encoding vectors between the principal components of the spatio-temporal part features of the scheduling requirement based on the kernel correlation performance analysis factor to obtain an interaction correlation matrix of the spatio-temporal part of the scheduling requirement as the interaction correlation representation information of the spatio-temporal part of the scheduling requirement. It should be understood that in the production scheduling scenario of a complex manufacturing system, the non-linear nature of the spatio-temporal coupling relationship makes it difficult for traditional encoding methods to directly support the spatio-temporal reasoning requirements of large language models (LLMs). Although spatio-temporal features are transformed into low-dimensional vectors through semantic embedding encoding, there are still multiple collinearity and high-order non-linear interaction characteristics in the potential correlation patterns between the time and space dimensions (such as the dynamic matching relationship between the equipment idle period and the material handling path). For example, the topological proximity of equipment coordinates may form a non-linear coupling with the overlap degree of its process time window. If only relying on independent time and space vector representations, it is easy for the LLM to ignore implicit correlation logics such as "sudden changes in buffer capacity caused by regional equipment collaboration" during the reasoning process. Based on this, in the technical solution of the present application, further spatio-temporal semantic correlation analysis is performed on the semantic embedding encoding vectors of the time-related part of the scheduling requirement and the semantic embedding encoding vectors of the space-related part of the scheduling requirement to obtain an interaction correlation matrix of the spatio-temporal part of the scheduling requirement. Specifically, the spatio-temporal semantic correlation analysis is a way of performing interactive parsing based on spatio-temporal performance correlation on the time-related part semantics and space-related part semantics of the scheduling requirement. In essence, it is a secondary deep decoupling and recombination of the spatio-temporal semantic embedding vectors of the scheduling requirement.

[0030] More specifically, first, perform principal component analysis on the semantic embedding coding vectors related to the scheduling demand time part and the semantic embedding coding vectors related to the scheduling demand space part respectively. Its core function is to strip redundant linearly correlated components from the original vectors. For example, the equipment idle time vector and the material waiting time vector may show strong linear correlation due to the production line beat synchronization. The principal component coding vectors extracted through orthogonal transformation can eliminate such collinearity interference and highlight independent influencing factors such as "the compression amount of the overlapping time of processes due to the space limitation of workstations". Next, use the principal component kernel correlation coding method to map the principal component vectors of the scheduling demand time-space part to a high-dimensional kernel space. Its essence is to learn the non-linear interaction mode between the time principal components (such as the phase offset of the process sequence) and the space principal components (such as the path crossing frequency between equipment) through a deep neural network. For example, in the kernel space, the principal component of the spatial accessibility of a certain equipment may form a high-order non-linear combination with its time window principal component, mapping out a potential conflict mode of "time window splitting caused by cross-regional handling". Further, the constructed kernel correlation performance analysis factor correlation topology matrix of the scheduling demand time-space part quantifies the above kernel space correlation relationship into a computable graph structure. By defining performance analysis factors strongly related to the scheduling goal (such as the "time overlap - space distance" synergy coefficient), the time-space coupling strength between equipment is converted into the weight value of the graph node edge. For example, when the overlap degree of the time idle windows of two pieces of equipment shows a high covariance with the reciprocal of their physical distance, the performance analysis factor will assign a higher numerical value to its edge weight, reflecting the risk of resource waste of "time matching but space isolation" between the two in the scheduling. Finally, through the message passing mechanism of the graph convolutional neural network (GCN), multi-level fusion of local time-space interactions (such as the space constraints of a single workstation) and global correlation patterns (such as the material flow congestion in the entire factory area) is carried out. In this way, not only can obvious correlation relationships such as "the gain offset of long-distance handling and parallel processes" be noticed, but also hidden correlation chains such as "path crossing causing AGV scheduling deadlocks" can be captured through high-order connections in the graph structure. When such refined correlation feature information is input into the LLM, the model can trace the root cause of time-space conflicts based on the topological propagation path. For example, it can be identified that the "abnormal idle period of equipment in a certain area" is actually a domino effect caused by the space limitation of the upstream workstation, thus realizing scheduling decisions that truly conform to the dynamic characteristics of complex manufacturing systems.

[0031] In an embodiment of the present application, the scheduling requirement kernel correlation information encoding subunit 1221 is configured to: perform feature principal component analysis on the semantic embedding encoding features of the time-related part of the scheduling requirement and the semantic embedding encoding vector features of the space-related part of the scheduling requirement respectively to obtain a set of semantic feature principal component encoding vectors of the time-related part of the scheduling requirement and a set of semantic feature principal component encoding vectors of the space-related part of the scheduling requirement; perform principal component kernel correlation encoding on each corresponding semantic feature principal component encoding vector of the time-related part of the scheduling requirement and the semantic feature principal component encoding vector of the space-related part of the scheduling requirement in the set of semantic feature principal component encoding vectors of the time-related part of the scheduling requirement and the set of semantic feature principal component encoding vectors of the space-related part of the scheduling requirement respectively to obtain a set of kernel correlation encoding vectors between the principal components of the spatio-temporal part of the scheduling requirement.

[0032] Specifically, performing feature principal component analysis on the semantic embedding encoding features of the time-related part of the scheduling requirement and the semantic embedding encoding vector features of the space-related part of the scheduling requirement respectively to obtain a set of semantic feature principal component encoding vectors of the time-related part of the scheduling requirement and a set of semantic feature principal component encoding vectors of the space-related part of the scheduling requirement, which is expressed by the scheduling requirement feature principal component analysis formula as:

[0033]

[0034]

[0035] Wherein, V1 is the semantic embedding encoding vector of the time-related part of the scheduling requirement, V2 is the semantic embedding encoding vector of the space-related part of the scheduling requirement, PCA(·) is feature principal component analysis, X is the set of semantic feature principal component encoding vectors of the time-related part of the scheduling requirement, Y is the set of semantic feature principal component encoding vectors of the space-related part of the scheduling requirement, x1, x2, x m are respectively the first, second, and m-th semantic feature principal component encoding vectors of the time-related part of the scheduling requirement in the set of semantic feature principal component encoding vectors of the time-related part of the scheduling requirement, y1, y2, y m are respectively the first, second, and m-th semantic feature principal component encoding vectors of the space-related part of the scheduling requirement in the set of semantic feature principal component encoding vectors of the space-related part of the scheduling requirement, Λ1 is the semantic feature diagonal matrix of the time-related part of the scheduling requirement, λ 11 and λ 1m are respectively the eigenvalues corresponding to x1 and x m Λ2 is the semantic feature diagonal matrix of the space-related part of the scheduling requirement, λ 21 and λ 2m are respectively the eigenvalues corresponding to y1 and y mThe corresponding eigenvalues. It should be understood that in the production scheduling scenario of complex manufacturing systems, due to the high dimensionality and heterogeneity of the original data, there are multicollinearity problems in the time-related and space-related semantic embedding features of scheduling requirements. There may be implicit linear correlations between time features (such as process timings and equipment idle windows) and space features (such as equipment coordinates and path topologies). For example, the physical proximity of equipment and the overlap of its idle periods may present redundant associations. This collinearity will lead to redundant feature expressions in subsequent correlation analysis, interfering with the model's extraction of the essential laws of spatio-temporal coupling relationships. Through feature principal component analysis (PCA), linearly correlated components can be stripped at the mathematical level. By deeply decoupling the time and space semantic embedding features, redundant linear association patterns in the feature spaces of the original scheduling requirement time-related part semantic embedding coding features and the scheduling requirement space-related part semantic embedding coding vectors can be eliminated. Specifically, through orthogonal transformation, high-dimensional features are projected onto the principal component space, so that time principal components (such as process phase offsets) and space principal components (such as path crossing frequencies) are transformed into linearly independent basis vectors. This transformation not only retains the statistical significance of the original features, but also highlights key influencing factors through the principle of variance maximization. For example, non-linear relationships such as "the time gain offset caused by long-distance transportation in parallel processes" are decomposed into independent principal components, providing a redundant-free and strongly interpretable feature basis for subsequent spatio-temporal interaction modeling. The set of principal component coding vectors of the scheduling requirement time-related part semantic features and the set of principal component coding vectors of the scheduling requirement space-related part semantic features generated after execution essentially construct a decoupled representation space for spatio-temporal features. Among them, the set of principal component coding vectors of the scheduling requirement time-related part semantic features can represent the dynamic phase features of process sequences (such as the compression amount of process overlap windows), and the set of principal component coding vectors of the scheduling requirement space-related part semantic features maps the topological characteristics of equipment layouts (such as the collaborative density of regional equipment). After eliminating the collinearity interference through the orthogonality of the principal components, they can more clearly reflect independent action mechanisms such as "time window offset caused by workbench space limitations". By adjusting the number of principal components retained, the feature dimension can be reduced while retaining key spatio-temporal coupling information, providing a lightweight but information-complete input for graph structure correlation coding, thereby improving the inference efficiency of subsequent large language models for implicit spatio-temporal conflicts.

[0036] Specifically, each corresponding principal component coding vector of the scheduling requirement time-related part semantic features and the principal component coding vector of the scheduling requirement space-related part semantic features in the set of principal component coding vectors of the scheduling requirement time-related part semantic features and the set of principal component coding vectors of the scheduling requirement space-related part semantic features are respectively subjected to principal component kernel correlation coding to obtain a set of kernel correlation coding vectors between the principal components of the scheduling requirement spatio-temporal part features, which is expressed by the principal component kernel correlation coding formula of the scheduling requirement as:

[0037]

[0038] where x i is the i-th scheduling requirement time-related partial semantic feature principal component coding vector in the set of scheduling requirement time-related partial semantic feature principal component coding vectors, y i is the i-th scheduling requirement space-related partial semantic feature principal component coding vector in the set of scheduling requirement space-related partial semantic feature principal component coding vectors, ‖·‖ represents the one-norm of a vector, α and β respectively represent trainable weighted hyperparameters, v iThe $i$-th kernel correlation encoding vector among the principal components of the spatio-temporal part features of scheduling requirements in the set of kernel correlation encoding vectors. It should be understood that after the principal component decomposition of the time and space semantic features, although the principal component encoding vectors of the time-related part semantic features and the space-related part semantic features of the scheduling requirements have eliminated linear redundancy, the non-linear interaction patterns in the spatio-temporal coupling relationship (such as the "synergistic effect of equipment spatial proximity and process time overlap") still cannot be directly characterized by linear principal components. Due to the complex non-linear correlations in the spatio-temporal constraints of the manufacturing system (for example, the asymmetric impact of long-distance transportation paths on the time windows of parallel processes), it is necessary to map to a high-dimensional kernel space through kernel correlation encoding to capture the implicit correlation features that are difficult to separate in the low-dimensional space. Specifically, by constructing a non-linear interaction channel between the spatio-temporal principal components, the time principal component (such as the process time sequence phase offset) and the space principal component (such as the equipment topology density) are re-associated in the high-dimensional kernel space. Among them, the principal component kernel correlation encoding uses a learnable non-linear transformation function to transform the orthogonalized principal component encoding vectors of the time-related part semantic features and the space-related part semantic features of the scheduling requirements into a joint representation that can reflect complex relationships such as "path crossing frequency and die change time compression ratio". This encoding mechanism essentially constructs a transition layer from the decoupled principal components to the topological features of the graph structure, enabling the high-order combination of the "spatial reachability principal component and the time window splitting principal component" to be mapped into an association vector that can be parsed by the graph convolutional network, providing an input basis with non-linear semantics for subsequent graph structure modeling. The generated set of kernel correlation encoding vectors among the principal components of the spatio-temporal part features of the scheduling requirements essentially forms a non-linear hyperplane representation of the spatio-temporal coupling relationship. Each kernel correlation encoding vector among the principal components of the spatio-temporal part features of the scheduling requirements quantifies the non-linear interaction intensity between the time principal component features (such as the distribution of equipment idle periods) and the space principal component features (such as the AGV path topology) as a geometric distance in the high-dimensional space. For example, the kernel space association vector between the "equipment cluster spatial density principal component" and the "process overlap time compression principal component" can represent the implicit law of "excessive regional equipment density causing process queuing delays". This encoding result enables the originally orthogonal principal component encoding vectors of the time-related part semantic features and the space-related part semantic features of the scheduling requirements to re-establish a dynamic association that conforms to the manufacturing logic in the graph structure space, providing a feature topology with physical significance for the subsequent graph convolutional neural network to identify the "spatio-temporal conflict chain caused by cross-station resource competition".

[0039] In an embodiment of the present application, the scheduling requirement association coding subunit 1222 is configured to: calculate the kernel association performance analysis factor between each two kernel association coding vectors among the kernel association coding vectors of the scheduling requirement spatio-temporal partial feature principal components to obtain a scheduling requirement spatio-temporal partial kernel association performance analysis factor association topology matrix; perform disorder optimization on the spatial distribution of the scheduling requirement spatio-temporal partial kernel for each kernel association coding vector among the kernel association coding vectors of the scheduling requirement spatio-temporal partial feature principal components to obtain an optimized set of kernel association coding vectors of the scheduling requirement spatio-temporal partial feature principal components; input the optimized set of kernel association coding vectors of the scheduling requirement spatio-temporal partial feature principal components and the scheduling requirement spatio-temporal partial kernel association performance analysis factor association topology matrix into a graph convolutional neural network model to obtain a scheduling requirement spatio-temporal partial interaction association matrix.

[0040] Specifically, calculating the kernel association performance analysis factor between each two kernel association coding vectors among the kernel association coding vectors of the scheduling requirement spatio-temporal partial feature principal components to obtain a scheduling requirement spatio-temporal partial kernel association performance analysis factor association topology matrix is expressed by the scheduling requirement kernel association performance analysis factor calculation formula as:

[0041]

[0042] where, v i,k and v j,k are respectively the eigenvalues at the k-th position in the i-th and j-th kernel association coding vectors among the kernel association coding vectors of the scheduling requirement spatio-temporal partial feature principal components, n is the number of eigenvalues in the kernel association coding vectors of the scheduling requirement spatio-temporal partial feature principal components, s(v i , v j ) is the scheduling requirement spatio-temporal partial kernel association performance analysis factor between v i and v j , M AFor the scheduling requirement, it is the correlation topology matrix of the kernel correlation performance analysis factors in the spatio-temporal part. It should be understood that in the embodiments of the present application, the kernel correlation performance analysis factor is a metric function, aiming to quantify the performance correlation or similarity degree between the kernel correlation coding vectors of the spatio-temporal part feature principal components of two scheduling requirements. The performance definition here has quite flexibility and task relevance, and needs to be customized according to specific application scenarios. For example, in some tasks, the performance can be understood as the synergistic effect of the kernel correlation coding vectors of the spatio-temporal part feature principal components of two scheduling requirements when predicting the target variable, while in the clustering task, it can refer to the similarity degree of the kernel correlation coding vectors of the spatio-temporal part feature principal components of two scheduling requirements in terms of data distribution. The selection of the kernel correlation performance analysis factor is directly related to the quality of the correlation topology matrix of the kernel correlation performance analysis factors in the spatio-temporal part of the scheduling requirement, and thus affects the learning effect of the subsequent graph neural network. After obtaining the kernel correlation performance analysis factor, it is constructed into the correlation topology matrix of the kernel correlation performance analysis factors in the spatio-temporal part of the scheduling requirement. The topology matrix is essentially an adjacency matrix, used to represent the connection relationship between nodes in the graph. In this context, the row and column indices of the correlation topology matrix of the kernel correlation performance analysis factors in the spatio-temporal part of the scheduling requirement correspond to the kernel correlation coding vectors between different spatio-temporal part feature principal components of the scheduling requirement, and the value of the element of the correlation topology matrix of the kernel correlation performance analysis factors in the spatio-temporal part of the scheduling requirement is the correlation strength calculated by the kernel correlation performance analysis factor. If the value of the kernel correlation performance analysis factor is higher, the value of the element of the correlation topology matrix of the kernel correlation performance analysis factors in the spatio-temporal part of the scheduling requirement is larger, and vice versa.

[0043] Specifically, perform disorder optimization on the spatial distribution of the spatio-temporal part of the scheduling requirement for each kernel correlation coding vector between the spatio-temporal part feature principal components of the scheduling requirement in the set of kernel correlation coding vectors between the spatio-temporal part feature principal components of the scheduling requirement to obtain the set of optimized kernel correlation coding vectors between the spatio-temporal part feature principal components of the scheduling requirement. It should be understood that although the set of kernel correlation coding vectors between the spatio-temporal part feature principal components of the scheduling requirement has captured the non-linear correlation between the spatio-temporal principal components, each kernel correlation coding vector v i The disorder of the spatial distribution under the action of the random potential field (such as the mapping misalignment between the device cluster topology and the process time window) will lead to the mismatch between the graph structure adjacency matrix (performance analysis factor topology) and the spatial embedding of the node features. For example, the kernel correlation coding vector v between the principal component of the device space density and the spatio-temporal part feature principal component of the process overlapping time scheduling requirement iIt may show a scattered distribution in the potential field space due to random initialization and cannot accurately reflect the actual coupling law of "regional device cooperation efficiency and process timing compression". This disorder will destroy the physical meaning of message passing in the graph convolution process. In this step, by constructing the initial scheduling requirement spatio-temporal partial feature cross-section function matrix M Bi As the spatial gauge field, the kernel correlation encoding vector v i of the main components of the spatio-temporal partial features of the scheduling requirement is subjected to compactification remapping, aiming to eliminate the geometric distortion caused by the random potential field. Specifically, each kernel correlation encoding vector v i of the main components of the spatio-temporal partial features of the scheduling requirement is multiplied by the initial scheduling requirement spatio-temporal partial feature cross-section function matrix M Bi to generate the compactified vector v Bi of the spatio-temporal partial features of the scheduling requirement cross-section, and the compactified matrix M C of the spatio-temporal partial features of the scheduling requirement cross-section is constructed through two-dimensional splicing, that is:

[0044]

[0045] M C =(v B1 T , v B2 T ,…)

[0046] where v i is the i-th kernel correlation encoding vector in the set of kernel correlation encoding vectors of the main components of the spatio-temporal partial features of the scheduling requirement, M Bi is the initial scheduling requirement spatio-temporal partial feature cross-section function matrix, is matrix multiplication, v Bi is the i-th compactified vector of the spatio-temporal partial features of the scheduling requirement cross-section in the sequence of compactified vectors of the spatio-temporal partial features of the scheduling requirement cross-section, (·, ·, …) is two-dimensional splicing processing, and M C is the compactified matrix of the spatio-temporal partial features of the scheduling requirement cross-section.

[0047] Subsequently, by minimizing the Gaussian correlation coefficient between the compactified matrix M C of the spatio-temporal partial features of the scheduling requirement cross-section and the correlation topology matrix M A of the spatio-temporal partial kernel correlation performance analysis factor of the scheduling requirement, the initial scheduling requirement spatio-temporal partial feature cross-section function matrix M Bi is iteratively corrected to obtain the optimized scheduling requirement spatio-temporal partial feature cross-section function matrix M Bi ′, that is:

[0048]

[0049] where, is position subtraction, ‖·‖ F is the F-norm of the matrix, σ 2 is M C and M A is the variance of the set composed of all matrix values of, exp is the value of the natural exponential function with the natural constant e as the base.

[0050] Thus, by further optimizing the spatio-temporal partial feature cross-section function matrix M Bi ′ to optimize the kernel correlation coding vector between the spatio-temporal partial features of the scheduling demand:

[0051]

[0052] where M Bi ′ is the spatio-temporal partial feature cross-section function matrix of the optimized scheduling demand, v i ′ is the optimized kernel correlation coding vector between the spatio-temporal partial features of the scheduling demand corresponding to the i-th kernel correlation coding vector between the spatio-temporal partial features of the scheduling demand.

[0053] This process is essentially a manifold calibration of the kernel correlation coding vector v i space between the spatio-temporal partial features of the scheduling demand, so that the correlation vector between the "equipment space reachability principal component" and the "process time compression principal component" is realigned to the true physical topological space of the manufacturing system under the constraint of the gauge field. For example, the scatter distribution is transformed into an ordered cluster reflecting the actual correlation density of "path crossing frequency - mold change time". The set of optimized kernel correlation coding vectors between the spatio-temporal partial features of the scheduling demand realizes an equivariant mapping from the potential field space to the physical topological space through gauge field correction. Each kernel correlation coding vector between the spatio-temporal partial features of the scheduling demand is constrained to a compact manifold isomorphic to the topological matrix of the performance analysis factor, so that the message passing of the graph convolutional neural network can be aggregated based on the true correlation strength. For example, an optimized kernel correlation coding vector between the spatio-temporal partial features of the scheduling demand can characterize the compact correlation strength between the "spatial principal component of the high-density equipment area" and the "parallel process time window principal component", accurately reflecting the physical law of "process delay accumulation caused by regional overload". This improvement in spatial distribution order enables the graph convolutional layer to more accurately capture the "domino-style spatio-temporal conflicts caused by sudden changes in buffer capacity", thereby enhancing the reasoning ability of the scheduling engine for implicit manufacturing logic.

[0054] Specifically, the set of optimized kernel correlation coding vectors between the spatio-temporal partial features of the scheduling demand and the topological matrix of the kernel correlation performance analysis factor of the spatio-temporal part of the scheduling demand are input into the graph convolutional neural network model to obtain the spatio-temporal partial interaction correlation matrix of the scheduling demand, which is expressed by the spatio-temporal partial interaction correlation formula of the scheduling demand as:

[0055]

[0056] Among them, GCM(·,·) represents the graph convolutional neural network model, and M c is the spatio-temporal partial interaction correlation matrix of scheduling requirements. It should be understood that in complex scheduling scenarios, spatio-temporal conflicts often manifest as implicit correlation chains of multi-hop propagation (such as the spatial limitation of a certain workstation causing queuing delays of multiple downstream processes). Through the hierarchical message passing mechanism of the graph convolutional neural network (GCN), the node-level spatio-temporal features (such as spatio-temporal constraints of a single device) and the graph structure-level correlation patterns (such as cross-regional resource competition network) need to be deeply fused to solve the modeling defects of traditional methods for long-range dependence relationships. Specifically, through the topological reasoning ability of the graph convolutional neural network, a multi-scale interaction representation of spatio-temporal features can be constructed. Among them, the optimized kernel correlation coding vector between the main components of the spatio-temporal partial features of the scheduling requirements is injected into the graph structure as node attributes (the adjacency matrix is defined by the correlation topological matrix of the kernel correlation performance analysis factors of the spatio-temporal part of the scheduling requirements), so that the GCN dynamically aggregates local features such as "collaborative idle time of spatially adjacent device groups" and "overlap degree of material flow paths across processes" during the message passing process, and captures high-order correlation patterns such as "global buffer capacity fluctuations caused by over-dense regional devices" through multi-layer convolution operations. This mechanism essentially realizes the evolution from atomic-level spatio-temporal constraints (single-node features) to system-level interaction laws (global graph semantics), providing an interpretable global spatio-temporal coupling matrix for the subsequent large language model, that is, the spatio-temporal partial interaction correlation matrix of scheduling requirements. By outputting the spatio-temporal partial interaction correlation matrix of scheduling requirements, this step deeply encodes the topological structure of the implicit spatio-temporal dependencies within the manufacturing system. Each element in the spatio-temporal partial interaction correlation matrix of scheduling requirements quantifies the dynamic influence strength between different combinations of spatio-temporal main components through weight learning of multi-layer graph convolution. For example, a certain row vector can represent the interaction weight between the "spatial density main component of device group A" and the "time window splitting main component of process chain B", reflecting the conflict intensity of "overload of devices in area A leading to an exponential increase in the waiting time of process B", and finally generating a scheduling plan that simultaneously meets local spatio-temporal constraints and global optimization goals.

[0057] Figure 5 It is a block diagram of a scheduling plan generation module in a complex manufacturing system production scheduling engine based on spatio-temporal reasoning of a large language model according to an embodiment of the present application. As Figure 5 shown, in the embodiment of the present application, the scheduling plan generation module 130 includes: a spatio-temporal partial interaction correlation adapter encoding unit 131, configured to input the spatio-temporal partial interaction correlation matrix of scheduling requirements into a large language model adapter to obtain a spatio-temporal scheduling requirement text sequence; a scheduling plan acquisition unit 132, configured to perform production scheduling based on the spatio-temporal scheduling requirement text sequence to obtain a scheduling plan.

[0058] Specifically, the scheduling requirement spatio-temporal partial interaction correlation adapter encoding unit 131 is used to input the scheduling requirement spatio-temporal partial interaction correlation matrix into the large language model adapter to obtain the scheduling spatio-temporal requirement text sequence. It should be understood that although the scheduling requirement spatio-temporal partial interaction correlation matrix has captured the spatio-temporal multi-dimensional coupling relationships (such as equipment collaboration and path conflict chains) in the scheduling requirements through the graph neural network, it is essentially still a high-dimensional numerical topological representation and cannot be directly aligned with the text understanding paradigm of the large language model (LLM). Therefore, the scheduling requirement spatio-temporal partial interaction correlation matrix is further input into the large language model adapter to obtain the scheduling spatio-temporal requirement text sequence. The core role of the large language model adapter is to build a bridge from the topological space to the semantic space. This adapter is essentially a feature-text mapping network based on the attention mechanism. By decoding the graph structure information in the scheduling requirement spatio-temporal partial interaction correlation matrix, it reconstructs it into a spatio-temporal semantic description sequence that can be parsed by the LLM, thus ensuring that the generated text sequence not only contains the quantitative relationships in the original data but also has domain semantic interpretability. Specifically, through the feature-text mapping network based on the attention mechanism, the adapter can dynamically assign weights to different elements in the scheduling requirement spatio-temporal partial interaction correlation matrix, highlighting the information that is most critical for the current task. For example, when analyzing the collaboration situation between devices, the adapter can adjust the attention weights to emphasize the device combinations with a higher degree of time window overlap or a shorter spatial distance, as these factors often have a significant impact on the scheduling decision. Among them, the adapter decomposes the scheduling requirement spatio-temporal partial interaction correlation matrix into multiple sub-matrices or feature vectors, and each sub-matrix or feature vector corresponds to a specific type of spatio-temporal relationship or the interaction between a certain group of devices. Then, for each sub-matrix or feature vector, a pre-trained feature-text mapping network is used for the mapping operation. This feature-text mapping network is trained on a large corpus related to manufacturing scheduling. It has mastered a certain degree of professional terms and expression methods and can generate corresponding descriptive texts according to the input features. For example, if a feature indicates that there is a tight time overlap and a close physical distance between two devices, the adapter may output a text fragment such as "Device A and Device B need to consider synchronous scheduling due to their close geographical location and compact process arrangement". It should be noted that during the entire conversion process, the adapter not only focuses on accurately translating numerical information into text descriptions but also endeavors to maintain the relationship coherence and logical consistency of the original data. In a specific embodiment of the present application, through context fusion, that is, when generating each descriptive text sentence, referring to the previously generated content and other related information. This helps to ensure that the finally formed scheduling spatio-temporal requirement text sequence constitutes a logically rigorous whole rather than a series of isolated sentences.

[0059] In an embodiment of the present application, a scheduling plan acquisition unit 132 is configured to input a scheduling basic information part and a scheduling spatio-temporal requirement text sequence into a production scheduling driver based on a large model to obtain a scheduling plan. It should be understood that in the production scheduling decision-making process of a complex manufacturing system, a globally optimized scheduling plan cannot be generated relying solely on the textual description of spatio-temporal constraints or isolated basic production parameters. The scheduling basic information part (such as order priority, product process route) is essentially static structured data, while the scheduling spatio-temporal requirement text sequence carries dynamic and coupled spatio-temporal semantic logic (for example, "the regional adjacency mold change time compression coefficient ≥ 0.7", "the path crossing frequency threshold triggers AGV rescheduling"). If the two are input into a large language model (LLM) separately, it is easy to cause semantic discontinuity during model inference. For example, the model may generate a compact schedule based on the order delivery date, but ignore potential conflicts such as "queue deadlocks caused by space limitations in the equipment cluster of high-priority orders". Furthermore, the scheduling basic information part and the scheduling spatio-temporal requirement text sequence are further input into a production scheduling driver based on a large model to obtain a scheduling plan. The core value of the production scheduling driver based on a large model lies in realizing collaborative reasoning with multi-modal input. The driver fuses the pre-set knowledge graph in the manufacturing field with the general reasoning ability of the LLM, transforms the scheduling basic information into an entity relationship graph such as equipment load and process chain, and at the same time interprets the spatio-temporal requirement text sequence into a computable constraint rule set. For example, when "Order A needs to be completed within 8 hours" in the basic information and "15% buffer time needs to be reserved for the collaborative operation of equipment group M1-M3" in the spatio-temporal text are input at the same time, the production scheduling driver will construct an associated mapping between the two in the implicit space - through the attention mechanism, it is recognized that equipment group M1-M3 is the core processing unit of Order A, and then the decision-making path of "the duration of non-critical processes needs to be compressed to meet the total delivery date" is derived. From the perspective of execution effect, this realizes the dual enhancement of domain knowledge and data-driven. On the one hand, the quantitative rules embedded in the spatio-temporal text sequence (such as "path crossing frequency > 2 times / hour triggers an alarm") activate the manufacturing scheduling common sense learned in the pre-training stage of the LLM, enabling the model to call patterned solutions such as "AGV avoidance strategy" and "mold change sequence optimization"; on the other hand, the production parameters in the basic information (such as equipment production capacity, bill of materials) provide physical boundary conditions for spatio-temporal constraints, avoiding the generation of "theoretically optimal but actually infeasible" scheduling plans. For example, when the text sequence indicates that "the handling distance in area 5 exceeds the limit", the driver will combine the number of AGVs and the load upper limit in the basic information, automatically rule out the simple strategy of "dispatching additional AGVs", and instead generate an innovative solution of "adjusting the process time sequence to stagger handling".

[0060] In summary, the production scheduling engine 100 of the complex manufacturing system based on large language model spatio-temporal reasoning according to the embodiments of the present application is elucidated. It obtains scheduling requirements and extracts time-related information and space-related information. Then, it jointly encodes the spatial features and the time constraint graph into spatio-temporal semantic vectors that can be parsed by the LLM, and explicitly models the multi-dimensional interaction relationship of "spatial proximity - time overlap degree - resource competition intensity" through the correlation analysis of the scheduling spatio-temporal semantics, enabling the LLM to achieve more reliable spatio-temporal causal reasoning, and finally generating a scheduling plan that meets complex spatio-temporal constraints. In this way, it can adapt to the high-dimensional spatio-temporal features in the manufacturing system, highlight the spatio-temporal information of the user's scheduling requirements to adapt to the input format of the LLM, and then focus on the non-linear spatio-temporal coupling relationship in the production scheduling of the complex manufacturing system, improving the generalization ability.

[0061] As described above, the production scheduling engine 100 of the complex manufacturing system based on large language model spatio-temporal reasoning according to the embodiments of the present application can be implemented in various terminal devices. In one example, the production scheduling engine 100 of the complex manufacturing system based on large language model spatio-temporal reasoning can be integrated into the terminal device as a software module and / or a hardware module. For example, the production scheduling engine 100 of the complex manufacturing system based on large language model spatio-temporal reasoning can be a software module in the operating system of the terminal device, or can be an application program developed for the terminal device; of course, the production scheduling engine 100 of the complex manufacturing system based on large language model spatio-temporal reasoning can also be one of the many hardware modules of the terminal device.

[0062] Alternatively, in another example, the production scheduling engine 100 of the complex manufacturing system based on large language model spatio-temporal reasoning and the terminal device can also be separate devices, and the production scheduling engine 100 of the complex manufacturing system based on large language model spatio-temporal reasoning can be connected to the terminal device through a wired and / or wireless network and transmit interaction information according to a predefined data format.

[0063] In several embodiments provided by the present invention, it should be understood that the disclosed devices, apparatuses and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For example, the division of modules is only a logical function division, and there can be other division methods in actual implementation.

[0064] The modules described as separate components may or may not be physically separated, and the components shown as modules may or may not be physical units, that is, they may be located in one place or distributed over multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in each embodiment of the present invention, the functional modules can be integrated in a processing unit, or each unit can exist physically alone, or two or more units can be integrated in one unit. The above integrated unit can be implemented in the form of hardware, or in the form of hardware plus software functional modules.

[0065] For those skilled in the art, it is obvious that the present invention is not limited to the details of the above exemplary embodiments, and without departing from the spirit or basic characteristics of the present invention, the present invention can be implemented in other specific forms.

[0066] Therefore, from any point of view, the embodiments should be regarded as exemplary and non-limiting. The scope of the present invention is defined by the appended claims rather than the above description. Therefore, all changes falling within the meaning and scope of the equivalent elements of the claims are intended to be included in the present invention. Any reference signs in the claims should not be regarded as limiting the claims involved.

[0067] In addition, it is obvious that the word "including" does not exclude other units or steps, and the singular does not exclude the plural. The multiple units or devices stated in the system claims can also be implemented by one unit or device through software or hardware. Words such as "second" are used to denote names and do not denote any particular order.

[0068] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not restrictive. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit of the technical solutions of the present invention.

Claims

1. A production scheduling engine for complex manufacturing systems based on spatio-temporal reasoning of large language models, characterized in that, Including: A scheduling requirement decomposition module, configured to obtain scheduling requirements input by a user, and extract a scheduling basic information part, a scheduling requirement time-related part, and a scheduling requirement space-related part from the scheduling requirements; A scheduling requirement spatio-temporal semantic interaction correlation encoding module, configured to perform scheduling spatio-temporal semantic correlation analysis based on a semantic feature space on the scheduling requirement time-related part and the scheduling requirement space-related part to obtain scheduling requirement spatio-temporal part interaction correlation representation information; A scheduling plan generation module, configured to determine a scheduling plan based on the scheduling requirement spatio-temporal part interaction correlation representation information and the scheduling basic information part, and by using a large language model; Wherein, the scheduling requirement spatio-temporal semantic interaction correlation encoding module is configured to: perform interactive parsing based on scheduling spatio-temporal performance correlation on the scheduling requirement time-related part and the scheduling requirement space-related part to obtain the scheduling requirement spatio-temporal part interaction correlation representation information.

2. The production scheduling engine of a complex manufacturing system based on spatio-temporal reasoning of large language models according to claim 1, wherein The scheduling requirement spatio-temporal semantic interaction correlation encoding module includes: A scheduling requirement spatio-temporal semantic embedding encoding unit, configured to perform semantic embedding encoding on the scheduling requirement time-related part and the scheduling requirement space-related part to obtain scheduling requirement time-related part semantic embedding encoding features and scheduling requirement space-related part semantic embedding encoding features; A scheduling requirement spatio-temporal part fine-grained interaction correlation unit, configured to perform fine-grained interaction analysis based on scheduling spatio-temporal deep semantic feature performance correlation on the scheduling requirement time-related part semantic embedding encoding features and the scheduling requirement space-related part semantic embedding encoding features to obtain the scheduling requirement spatio-temporal part interaction correlation representation information.

3. The production scheduling engine of a complex manufacturing system based on spatio-temporal reasoning of a large language model according to claim 2, wherein The scheduling requirement spatio-temporal part fine-grained interaction correlation unit includes: A scheduling requirement kernel correlation information encoding sub-unit, configured to perform kernel correlation information encoding based on principal component analysis on the scheduling requirement time-related part semantic embedding encoding features and the scheduling requirement space-related part semantic embedding encoding vector features to obtain a set of kernel correlation encoding vectors between principal components of scheduling requirement spatio-temporal part features; A scheduling requirement correlation encoding sub-unit, configured to construct a kernel correlation performance analysis factor between every two kernel correlation encoding vectors between principal components of scheduling requirement spatio-temporal part features in the set of kernel correlation encoding vectors between principal components of scheduling requirement spatio-temporal part features, and perform graph structure-based correlation encoding on the set of kernel correlation encoding vectors between principal components of scheduling requirement spatio-temporal part features based on the kernel correlation performance analysis factor to obtain a scheduling requirement spatio-temporal part interaction correlation matrix as the scheduling requirement spatio-temporal part interaction correlation representation information.

4. The production scheduling engine of a complex manufacturing system based on spatio-temporal reasoning of a large language model according to claim 3, wherein The scheduling requirement kernel correlation information encoding sub-unit is configured to: Respectively perform feature principal component analysis on the scheduling requirement time-related part semantic embedding encoding features and the scheduling requirement space-related part semantic embedding encoding vector features to obtain a set of scheduling requirement time-related part semantic feature principal component encoding vectors and a set of scheduling requirement space-related part semantic feature principal component encoding vectors; Perform principal component kernel association encoding on each corresponding pair of the semantic feature principal component encoding vectors of the scheduling requirement time-related part and the semantic feature principal component encoding vectors of the scheduling requirement space-related part in the set of semantic feature principal component encoding vectors of the scheduling requirement time-related part and the set of semantic feature principal component encoding vectors of the scheduling requirement space-related part to obtain a set of kernel association encoding vectors between the principal components of the scheduling requirement spatio-temporal part features.

5. The production scheduling engine of a complex manufacturing system based on spatio-temporal reasoning of a large language model according to claim 4, characterized in that, The scheduling requirement association encoding subunit is used for: Calculate the kernel association performance analysis factor between every two kernel association encoding vectors between the principal components of the scheduling requirement spatio-temporal part features in the set of kernel association encoding vectors between the principal components of the scheduling requirement spatio-temporal part features to obtain a kernel association performance analysis factor association topology matrix for the scheduling requirement spatio-temporal part; Perform disorder optimization of the kernel space distribution of the scheduling requirement spatio-temporal part on each kernel association encoding vector between the principal components of the scheduling requirement spatio-temporal part features in the set of kernel association encoding vectors between the principal components of the scheduling requirement spatio-temporal part features to obtain an optimized set of kernel association encoding vectors between the principal components of the scheduling requirement spatio-temporal part features; Input the optimized set of kernel association encoding vectors between the principal components of the scheduling requirement spatio-temporal part features and the kernel association performance analysis factor association topology matrix of the scheduling requirement spatio-temporal part into a graph convolutional neural network model to obtain the interaction association matrix of the scheduling requirement spatio-temporal part.

6. The production scheduling engine of a complex manufacturing system based on spatio-temporal reasoning of a large language model according to claim 5, characterized in that, The scheduling plan generation module includes: A scheduling requirement spatio-temporal part interaction association adapter encoding unit, which is used to input the interaction association matrix of the scheduling requirement spatio-temporal part into a large language model adapter to obtain a scheduling spatio-temporal requirement text sequence; A scheduling plan acquisition unit, which is used to perform production scheduling based on the scheduling spatio-temporal requirement text sequence to obtain a scheduling plan.

7. The production scheduling engine of a complex manufacturing system based on spatio-temporal reasoning of a large language model according to claim 6, characterized in that The scheduling plan acquisition unit is used for: Inputting the scheduling basic information part and the scheduling spatio-temporal requirement text sequence into a production scheduling driver based on a large model to obtain a scheduling plan.

Images