Intelligent supply chain scheduling method and system of a large model fusion

Abstract

This invention provides an intelligent supply chain scheduling method and system that integrates a large-scale model, belonging to the field of intelligent supply chain scheduling technology. The method includes first collecting production line sensor signals, supplier engineering change notifications, and market commentary text. Implicit quality characteristics of component batches are extracted by analyzing the signals, and a technology transmission network with the impact of changes as weights is constructed based on this. User sentiment in the market commentary is quantified as tolerance constraints on specific technical parameters and injected back into the network to form a constraint network. Supply chain disturbance events are simulated on this constraint network, and the strength of their transmission path's impact on performance constraint boundaries is evaluated, thereby generating preventative scheduling plans. Finally, with the goal of minimizing inter-stage switching costs, a full-cycle, staged scheduling instruction sequence is optimized. This invention achieves a closed loop from market perception to preventative mobilization, enhancing the supply chain's proactive immunity and overall resilience to internal and external uncertainties.

Description

Technical Field

[0001] This invention relates to the field of intelligent supply chain scheduling technology, and more specifically, to an intelligent supply chain scheduling method and system that integrates a large model. Background Technology

[0002] In the field of intelligent supply chain scheduling, the combination of big data and predictive models has significantly improved the efficiency of traditional planning and execution. However, existing technologies still face severe challenges when dealing with the highly dynamic and complex scenario of consumer electronics product launches, which involves multiple sources of disturbances. This scenario not only involves massive amounts of heterogeneous and rapidly evolving data from multiple sources, such as production line sensors, engineering changes from multiple tiers of suppliers, and feedback from the open market, but also requires the scheduling system to proactively absorb the dual uncertainties caused by technological changes and market sentiment in order to ensure the stable delivery of new products within a strict timeframe.

[0003] Current mainstream solutions primarily focus on the analysis of structured historical data and the optimization of static rules, or place the aforementioned multi-source data in independent processing flows. This leads to insufficient assessment of the cascading impact of engineering changes in the manufacturing chain, a severe lag in responding to dynamic market feedback, and difficulty in effectively transforming unstructured market sentiment into preventative supply chain constraints. Its core flaw lies in the lack of deep semantic connections and closed-loop feedback between various data and decision-making modules, making the final scheduling strategy often passive and rigid in response to sudden disturbances, failing to achieve an adaptive balance between efficiency and resilience.

[0004] Therefore, there is an urgent need for an intelligent supply chain scheduling method and system that integrates large models to solve the above-mentioned technical problems. Summary of the Invention

[0005] The purpose of this invention is to provide an intelligent supply chain scheduling method and system that integrates large-scale models to improve the aforementioned problems. To achieve this objective, the technical solution adopted by this invention is as follows:

[0006] Firstly, this application provides an intelligent supply chain scheduling method that integrates large-scale models, including:

[0007] Acquire timing sensor signal streams recording component testing processes on assembly and testing production lines, original documents of batch engineering change notifications from all component suppliers, and unstructured text streams of competitor feature experience and fault feedback from social media platforms;

[0008] Based on the time-series sensor signal stream, implicit feature extraction and characterization processing of component batch quality are performed to construct implicit feature vectors representing components in each batch.

[0009] Based on the original change notification document and the implicit feature vector of the component, an adaptive construction process of the change impact network is performed. The technology impact network that describes the transmission relationship of the change in the manufacturing chain is generated with the component and process parameters as nodes and the offset as the edge weight.

[0010] The unstructured text stream data and the technology influence network are subjected to inverse mapping of market demand disturbances and technology constraints. The market demand disturbances after inverse mapping are then coupled back to the technology influence network as new constraints, resulting in a constraint network that reshapes the boundaries of network nodes.

[0011] Based on the aforementioned technology, supply chain disturbance simulations are performed on the network, and the degree of disruption to the boundary conditions of nodes in the constrained network caused by the simulated transmission paths is calculated to generate preventive scheduling plans.

[0012] Based on the aforementioned preventive scheduling plan, a multi-stage decision sequence optimization coupled with the product release cycle is performed, wherein the scheduling instruction sequence is selected with the goal of minimizing the total switching cost and meeting the network resilience threshold of each stage.

[0013] Secondly, this application also provides an intelligent supply chain scheduling system that integrates large-scale models, including:

[0014] The acquisition unit is used to acquire timing sensor signal streams that record component testing processes on the assembly and testing production line, original documents of batch engineering change notices from all component suppliers, and unstructured text stream data from social media platforms regarding competitor feature experience and fault feedback.

[0015] The extraction unit is used to extract and characterize the implicit features of component batch quality based on the time-series sensor signal stream, and to construct an implicit feature vector characterizing each batch of components.

[0016] The construction unit is used to adaptively construct the change impact network based on the original change notification document and the implicit feature vector of the component, and generate a technology impact network that describes the transmission relationship of the change in the manufacturing chain, with the component and process parameters as nodes and the offset as the edge weight.

[0017] The mapping unit is used to perform reverse mapping processing of market demand disturbances and technical constraints based on the unstructured text stream data and the technology influence network, and to back-couple the reverse-mapped market demand disturbances as new constraints to the technology influence network to obtain a constraint network that reshapes the boundaries of network nodes.

[0018] The simulation unit is used to simulate supply chain disturbances based on the technology's impact on the network, calculate the degree of damage to the boundary conditions of nodes in the constrained network caused by the simulated transmission path, and generate preventive scheduling plans.

[0019] The optimization unit is used to perform multi-stage decision sequence optimization coupled with the product release cycle according to the preventive scheduling plan, wherein the scheduling instruction sequence is selected with the goal of minimizing the total switching cost and meeting the network resilience threshold of each stage.

[0020] The beneficial effects of this invention are as follows:

[0021] This invention acquires multi-source heterogeneous data from production lines, suppliers, and the market, and inversely transforms unstructured market sentiment into supply chain technical constraints, constructing a dynamic network model that integrates engineering technology influences and market perceptions. Based on this, it simulates the transmission effects of potential disturbances within the network and assesses their impact on performance boundaries defined by market demands, thereby generating preventative scheduling plans that proactively block risk transmission. Finally, these plans are combined with the product launch cycle for multi-stage optimization, resulting in a fixed sequence of scheduling instructions with the lowest total switching cost while meeting resilience requirements at each stage. This invention achieves a shift from passive response to proactive immunity, significantly enhancing the overall resilience, adaptability, and delivery assurance level of the supply chain in the face of internal changes and external market uncertainties.

[0022] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing embodiments of the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description, claims, and drawings. Attached Figure Description

[0023] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram of the intelligent supply chain scheduling method based on the fusion of large models as described in this embodiment of the invention;

[0025] Figure 2 This is a schematic diagram of the intelligent supply chain scheduling system structure based on the integrated large model described in this embodiment of the invention.

[0026] In the figure: 701, acquisition unit; 702, extraction unit; 703, construction unit; 704, mapping unit; 705, simulation unit; 706, optimization unit. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0028] It should be noted that similar reference numerals and letters in the following figures indicate similar items; therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures. Furthermore, in the description of this invention, terms such as "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0029] Example 1:

[0030] This embodiment provides an intelligent supply chain scheduling method that integrates large models.

[0031] See Figure 1 The figure shows that the method includes steps S1, S2, S3, S4, S5 and S6.

[0032] Step S1: Obtain the timing sensor signal streams that record the component testing process on the assembly and testing production line, the original documents of batch engineering change notices from all component suppliers, and the unstructured text stream data on competitor feature experience and fault feedback from social media platforms;

[0033] This step understandably involves acquiring raw, heterogeneous, and unprocessed data from three key dimensions: the manufacturing site, the upstream supply chain, and the market end. Specifically, the timing sensor signal streams from the assembly and testing lines include real-time recordings of various continuous signals such as voltage, current, timing, images, and acoustic vibrations generated during component insertion, soldering, functional testing, and aging testing. These signals are typically captured directly by high-precision sensors and data acquisition units deployed on the production line. The original documents of batch engineering change notifications from all component suppliers refer to engineering change orders and technical specification revision records transmitted in unstructured or semi-structured form. These documents directly describe changes in component specifications, materials, processes, or supply status, and are obtained through enterprise resource planning (ERP) or supply chain collaboration platform interfaces. Unstructured text stream data from social media platforms regarding competitor product features and fault feedback encompasses user comments on public forums, review websites, and social media. These text streams are collected and aggregated in real-time or near real-time using authorized web crawling technology, capturing collective market skepticism and potential quality trends.

[0034] Step S2: Based on the time-series sensor signal stream, perform implicit feature extraction and characterization processing of component batch quality to construct an implicit feature vector characterizing each batch of components.

[0035] Understandably, this step processes the sensor signal stream from the production line. Its purpose is to penetrate the surface-level test results and extract deeper, implicit characteristics from continuous time-series signals that characterize the inherent quality consistency and reliability of component batches. This goes beyond traditional yield statistics; by capturing subtle pattern differences within the signals, it encodes the overall performance and stability of each batch of components into a computable mathematical vector, thus laying the foundation for subsequent quantitative assessments of the impact of engineering changes on specific batches. In this step, step S2 includes steps S21, S22, and S23.

[0036] Step S21: Perform multi-process signal coupling and decomposition processing based on the time sequence sensor signal stream. By aligning and decomposing the sensor signals of different test processes in the time and frequency domain, separate the mixed signal components that characterize the inherent properties of the components, the contact state of the fixture, and the environmental noise to obtain the decoupled component body response signal sequence.

[0037] Understandably, this step involves multi-stage signal coupling and decomposition of the timing sensor signal streams from multiple testing processes. By aligning and analyzing these mixed signals in the time and frequency dimensions, different signal components that characterize the inherent electrical characteristics of the components themselves, the contact impedance changes of the test fixtures, and the environmental background noise are effectively separated. This results in a pure decoupled signal sequence that reflects only the performance response of the components themselves. This step eliminates the interference of fixtures and environmental variables in the manufacturing process on quality assessment, ensuring that subsequent analysis is based on the actual state of the components.

[0038] First, the sensor signals from all processes are timestamped and resampled to ensure strict alignment of the data on the time axis. Next, a short-time Fourier transform is performed on each synchronized signal to convert it from the time domain to the time-frequency domain, resulting in a complex matrix where rows correspond to frequency points and columns to time windows. The complex observations from all processes at the same time window and frequency point are combined into an observation vector, thus transforming the multi-process signal coupling problem into a blind source separation problem. In this scenario, three source signals are pre-defined: component body response, fixture contact noise, and environmental background noise. A demixing matrix is ​​solved using the FastICA algorithm, which iteratively optimizes the demixing matrix by maximizing the non-Gaussianity of the output components. Once the algorithm converges, the estimated source signals can be separated from the mixed observation signals using the demixing matrix. We select the component with the most stable time-domain waveform and the highest correlation with product specifications as the decoupled component body response signal sequence output.

[0039] The formulas for multi-process signal coupling and decomposition are shown below:

[0040] ;

[0041] in, In time The observed mixed signal vector from all processes has a dimension of , for The mixture matrix, It is a three-dimensional vector, containing time. The three independent source signals correspond to the component body response, fixture contact noise, and ambient background noise, respectively.

[0042] Step S22: Perform cross-batch abnormal pattern clustering and tracing processing based on the component body response signal sequence, and associate the processed abnormal patterns with the supplier production batch identifier to obtain an abnormal pattern feature cluster with supplier batch tracing label;

[0043] Understandably, this step identifies recurring fluctuations or morphological patterns that deviate from the normal range in the signal sequence, categorizes and groups these similar abnormal patterns, and precisely associates each pattern cluster with a specific supplier and its production batch identifier, thereby forming an abnormal pattern feature cluster with a clear source label. This allows previously isolated test anomalies to be traced back to a specific production batch of a specific supplier, thus achieving the root cause location of quality problems.

[0044] This step first extracts a set of time-domain and frequency-domain features from the signal sequence of each component, including the mean, standard deviation, skewness, kurtosis, and the first few moments of the spectrum. Then, based on data from historical "normal" batches, a multivariate Gaussian distribution model of these features is calculated, namely the mean vector and covariance matrix. For each component in the new batch, the Mahalanobis distance from its feature vector to this "normal" distribution is calculated to quantify its anomalousness. If this distance exceeds a preset threshold (usually set to 3.0, corresponding to approximately 99.7% confidence interval), the component is marked as an anomalous sample. Subsequently, the feature vectors of all anomalous samples from all batches are collected and grouped using the K-means clustering algorithm. K-means iteratively optimizes the data into K clusters to minimize the sum of squared Euclidean distances from each sample to the center of its cluster. After clustering, each cluster represents a specific anomalous pattern.

[0045] The formula for calculating the Mahalanobis distance in anomaly detection is as follows:

[0046] ;

[0047] in, The Mahalanobis distance for anomaly detection. These are the feature vectors extracted from the component signals. This is a vector of feature mean values ​​calculated based on historical normal batch data. For transpose, Given the covariance matrix of historical normal batch data, if A value greater than 3 is considered abnormal.

[0048] The objective function of the K-means algorithm is shown below:

[0049] ;

[0050] in, Let K-means be the objective function. The preset number of clusters, For the first The sample set of each cluster It is the first The center vector of each cluster Let be the feature vector of a certain abnormal sample. It is the square of the Euclidean distance.

[0051] Step S23: Based on the abnormal pattern feature cluster with supplier batch traceability label, perform process robustness implicit coding vector synthesis processing to obtain implicit feature vector characterizing the overall quality of the batch of components.

[0052] It is understandable that this step uses anomaly pattern feature clusters with batch traceability tags to synthesize encoded vectors. By analyzing the statistical distribution, frequency of occurrence, and correlation with key process parameter settings of the patterns in the feature clusters, a compact mathematical vector is generated. This vector not only summarizes the overall quality level of the components in the batch during testing, but also implies its tolerance to fluctuations in the production process. This provides a deep and computable characteristic representation for the subsequent quantitative analysis of the impact of engineering changes on batch quality.

[0053] This step uses a pre-trained autoencoder neural network to process the dataset. An autoencoder is an unsupervised neural network consisting of an encoder and a decoder, aiming to learn an efficient representation of the input data. During training, a large amount of historical anomaly feature data is used to optimize the network parameters by minimizing the reconstruction error between the input features and the decoder output. After training, each feature vector from a pre-defined batch of anomaly feature sets is input into the encoder to obtain a corresponding low-dimensional encoded vector. Then, we take the arithmetic mean of these low-dimensional encoded vectors to obtain a fixed-dimensional vector.

[0054] The reconstruction error loss function of the autoencoder is shown below:

[0055] ;

[0056] in, Let the reconstruction error loss function be the autoencoder. The total number of training samples, For parameters encoder network, This is historical anomalous feature data. For parameters decoder network, For each feature vector in the pre-defined batch of abnormal feature sets, For serial number, For the first Anomaly feature vectors used for training

[0057] The formula for calculating the implicit feature vector of the synthesized batch is as follows:

[0058] ;

[0059] in, For the calculated batch The implicit eigenvectors. For batch The set of feature vectors of all abnormal samples For batch The number of feature vector sets of all abnormal samples These are the feature vectors extracted from the component signals.

[0060] Step S3: Based on the original change notification document and the implicit feature vector of the component, perform adaptive construction processing of the change impact network, and use the component and process parameters as nodes and the offset as the edge weight to generate a technology impact network that describes the transmission relationship of the change in the manufacturing chain.

[0061] Understandably, this step combines static engineering change documents with dynamic, quantified batch quality characteristics to construct a dynamic transmission network. This process simulates the chain reaction that changes may trigger in the manufacturing chain. By calculating the expected offset of the change to the implicit feature vector of components, and using the offset as weight, a directed, weighted influence relationship is established between components and process parameters. This transforms a point-like change event into a network topology that may affect multiple manufacturing stages, providing a model basis for proactively assessing the ripple effect of changes. In this step, step S3 includes steps S31, S32, and S33.

[0062] Step S31: Extract the parameters associated with the change items and affected batches based on the original document of the change notification and the implicit feature vector, and map the extracted associated parameters to the corresponding implicit feature vectors of the components to obtain the set of change parameters and their associated batches and feature vectors.

[0063] Understandably, this step first parses the specific changes from the original engineering change notification document, including modifications to component models, specifications, or manufacturing processes, and identifies the specific component production batches affected by these changes. Then, these extracted change parameters are associated and mapped with the implicit feature vectors of the corresponding batches of components, thereby obtaining a structured set of change parameters, which clearly records the specific batch associated with each change and its corresponding quality characteristic representation.

[0064] The scoring function for joint extraction of entity relations is as follows:

[0065] ;

[0066] in, For the input sequence and specific entity label sequences Overall score Given the length of the input text sequence, This is the trainable weight matrix for the entity recognition part. For the first in the input sequence The hidden state vector of each word output from the penultimate layer of the model. For all possible entity pairs in the text, The sequence number of the entity pair. This is the trainable weight matrix for the relation classification part. To make the first The word and the first The hidden state vectors of each word are concatenated.

[0067] Step S32: Based on the changed parameter set and its associated batch and feature vector, perform feature offset calculation processing caused by the change. By simulating the changed parameter values ​​and inputting them into the feature vector generation model, calculate the perturbation direction and magnitude of the original associated implicit feature vector to obtain the quantified feature offset vector.

[0068] It is understandable that this step calculates the feature offset caused by the change in the parameter set. By inputting the simulated changed parameter values ​​into the same model previously used to generate implicit feature vectors, the direction and magnitude of the disturbance of the original feature vectors in multiple quality dimensions are calculated. Finally, a quantified feature offset vector is output, which accurately describes the expected impact of the engineering change on the intrinsic quality characteristics of a specific batch of components.

[0069] The Jacobian matrix of the implicit feature generation model is shown below:

[0070] ;

[0071] in, The calculated feature offset vector has dimensions of... , Jacobian matrix The List, For the first The amount of change in each input parameter.

[0072] Step S33: Construct a directed weighted network based on the quantized feature offset vector. Using component batches and manufacturing process parameters as nodes, and the disturbance direction and magnitude represented by the feature offset vector as the weight and direction of the edges, construct a technology influence network for the multi-level transmission relationship of change notifications in the manufacturing chain.

[0073] Understandably, this step utilizes quantized feature offset vectors to construct a directed weighted network, where nodes represent affected component batches and related manufacturing process parameters, and the weights and directions of directed edges are determined by the disturbance intensity and direction defined by the feature offset vectors. This forms a technology impact network that can describe and quantify the multi-level chain transmission effects that a single engineering change may trigger in the manufacturing chain through process and material dependencies, providing a model basis for dynamically assessing the diffusion risk of changes.

[0074] The formula for calculating the network edge weights influenced by technology is shown below:

[0075] ;

[0076] in, For parameter nodes Point to batch node The weight of the edge. For the parameter node When changes occur, for a specific batch The L2 norm of the obtained feature offset vector is calculated.

[0077] Step S4: Perform reverse mapping processing of market demand disturbance and technology constraint based on the unstructured text stream data and the technology influence network, and couple the reverse-mapped market demand disturbance as a new constraint condition back to the technology influence network to obtain a constraint network that reshapes the boundaries of network nodes.

[0078] Understandably, this step involves the market's collective sentiment regarding competitors' defects, which, through semantic analysis and correlation mapping, is decoded inversely into quantified tolerance boundaries for specific technical parameters of their own products (operating temperature range, load response time). These constraints, derived from market sentiment, are injected back into the technology influence network, dynamically reshaping the stability boundaries of key nodes in the network and directly linking the supply chain's resilience targets to the market's actual tolerance. In this step, step S4 includes steps S41, S42, and S43.

[0079] Step S41: Perform market sentiment quantification processing on the unstructured text stream data for competitor functional defects. By performing fine-grained sentiment and entity analysis on the text stream, identify and quantify the intensity of collective dissatisfaction expressed by users regarding the performance defects of competitors, and obtain a market sentiment intensity vector.

[0080] Understandably, this step first quantifies market sentiment regarding competitor functional defects by processing unstructured text stream data obtained from social media and other channels. Through fine-grained sentiment analysis of massive amounts of user comments and identification of specific performance defect entities mentioned within them, subjective emotions such as user complaints and dissatisfaction with problems exposed by competitors in specific functions and scenarios are aggregated and calculated into quantifiable group sentiment intensity indicators for different performance dimensions, thereby obtaining a structured market sentiment intensity vector. This process transforms vague market reputation into calculable data.

[0081] This step begins with preprocessing the text stream, including noise removal, word segmentation, and part-of-speech tagging. The core processing employs a dual-task deep learning model based on an attention mechanism: the entity recognition task detects "performance defect entities" mentioned in the text (such as "short battery life," "severe overheating," and "screen flickering"); the sentiment analysis task performs fine-grained sentiment classification on sentences or contexts containing these entities, outputting sentiment polarity (positive, negative, neutral) and intensity values. For each identified "performance defect entity," the system uses topic modeling (LDA) to cluster its context, merging different but essentially similar defects into several "core performance issues" categories. Subsequently, for each core performance issue, the system aggregates sentiment intensity within a sliding time window. The intensity calculation comprehensively considers three factors: the frequency of the issue's mention in the text, the average intensity of negative sentiment accompanying its mention, and the central tendency of negative evaluations. Finally, a scalarized "market sentiment intensity" value is output for each core performance issue, and the intensity values ​​of all core performance issues are arranged in a fixed order to form a "market sentiment intensity vector."

[0082] The formula for aggregating market sentiment intensity is shown below:

[0083] ;

[0084] in, For core performance issues Market sentiment intensity value This represents the total number of discrete time intervals within the time window used for aggregation calculations. For time, , and The weighting coefficients for the three factors—equilibrium frequency, average intensity, and concentration—are 0.4, 0.4, and 0.2, respectively. In time The text stream mentions core performance issues. Number of times, In time All the questions mentioned The average intensity of negative emotions accompanying the text. In time Mentioning the issue The standard deviation of the intensity of negative emotions.

[0085] Step S42: Based on the market sentiment intensity vector and the technology influence network, perform the mapping and constraint transformation of sentiment vector to network technology nodes. By converting the market sentiment intensity vector into a tolerance threshold for the fluctuation range of relevant technology node parameters, a set of performance constraint conditions for the corresponding network nodes is obtained.

[0086] Understandably, this step combines the market sentiment intensity vector obtained in the previous step with the technology influence network, and performs mapping and constraint transformation of the sentiment vector to network technology nodes. Through preset association rules, it maps the specific performance problem pointed to by market sentiment to the component batch node or process parameter node in the technology influence network that directly affects the performance, and transforms the sentiment intensity into a more stringent tolerance threshold for the allowable fluctuation range of the node parameter, thereby generating a set of performance constraints for a specific network node, realizing the transformation from subjective market feedback to engineering and technical executable tolerance requirements.

[0087] For each intensity value in the emotion vector, the system finds the set of associated nodes based on the historical database. Emotion intensity is translated into a "tightening requirement" for the tolerance range of the node's parameters. Its tolerance threshold is adjusted by a monotonically decreasing function based on the emotion intensity of all associated questions; the stronger the emotion, the tighter the tolerance range.

[0088] The function for calculating the tightening ratio of the node parameter tolerance range is shown below:

[0089] ;

[0090] in, To influence nodes in the network through technology The parameter tolerance range adjustment ratio, This is the lower limit of the minimum allowable adjustment ratio. The global sensitivity coefficient, adjusted for emotional intensity to tolerance, is set to 0.3. To link nodes through historical database The collection of all core performance issues, For standard sigmoid functions, For core performance issues Market sentiment intensity value This serves as the baseline value for the threshold of emotional intensity.

[0091] Step S43: Perform reverse injection and boundary reshaping of the constraint conditions on the technology influence network according to the performance constraint condition set. By applying the tolerance threshold in the performance constraint condition set as a new boundary condition to the corresponding parameter node in the technology influence network, the feasible domain of the node state is modified, and the constraint network after the node boundary is reshaped is obtained.

[0092] It is understandable that this step uses the new parameter tolerance threshold specified by each constraint in the set as a hard or soft boundary, and assigns it to the corresponding parameter node in the technology influence network, thereby shrinking or modifying the allowable value range of the node's state, i.e. its feasible region. When the feasible region of one or more nodes in the network is modified, the change transmission relationship and risk path represented by the entire network are also dynamically updated, thus obtaining a constraint network that incorporates the external market tolerance requirements and whose node boundaries have been reshaped.

[0093] It is understandable that in a technology-influenced network, the state (i.e., parameter value) of each parameter node is considered to be variable within a continuous interval, which is its feasible region. The processing directly assigns the new tolerance threshold given in the constraint set to the corresponding node as its new, more stringent feasible region boundary.

[0094] The formula for reshaping the feasible domain of network nodes due to technological influence is shown below:

[0095] ;

[0096] in, To influence nodes in the network through technology The feasible region is the set of all possible values ​​that its parameters can take. For nodes The variable values ​​of the parameters they represent. For the set of real numbers, For nodes The new lower bound for the parameters is given by the set of performance constraints. For nodes The new upper bound for the parameters is given by the set of performance constraints.

[0097] Step S5: Simulate supply chain disturbances based on the technology-affected network, calculate the degree of damage to the boundary conditions of nodes in the constrained network by the simulated transmission path, and generate a preventive scheduling plan.

[0098] Understandably, this step utilizes a pre-constructed network model incorporating market constraints to conduct proactive risk stress testing. This step simulates various typical supply chain disruption events within the technology impact network, such as shortages of critical materials or batch quality mutations, and observes their propagation paths along the network. More importantly, it assesses the impact and disruption of these simulated propagation paths on node boundary conditions representing market resilience, thereby identifying vulnerable links and high-risk paths in the supply chain, providing precise quantitative evidence for generating targeted preventative scheduling measures. In this step, step S5 includes steps S51, S52, and S53.

[0099] Step S51: Based on the technology impact network, simulate disturbance events of typical supply chain risks, wherein a node state anomaly representing a shortage of key components, a sudden change in batch quality, or an emergency engineering change is triggered with a preset probability, and the process of the anomaly state spreading along the transmission relationship defined by the network edge weights is simulated to obtain a set of simulated events including the initial disturbance node and the subsequent transmission path.

[0100] Understandably, this step first simulates disturbance events of typical supply chain risks based on the technology impact network. Specifically, it simulates the triggering of abnormal states such as supply interruptions, quality parameter drift, or emergency engineering changes at network nodes representing key components, specific production batches, or process parameters using statistical probability. Based on the causal and impact strength relationships defined by the edge weights in the network, the abnormal state is allowed to propagate and pass up and down the network connection to upstream and downstream nodes. Through a large number of such Monte Carlo simulations, a set of simulated events containing different initial disturbance sources, different transmission paths, and different impact ranges is obtained, thereby systematically revealing the potential vulnerable links in the supply chain network.

[0101] First, based on the node type (e.g., "critical components," "ordinary batches," "core process parameters") and a preset trigger probability, a group of nodes is randomly selected as the initial disturbance nodes, and each is assigned an initial state offset (e.g., a supply shortage node has a state offset of 1, indicating complete failure; a quality mutation node has a state offset of 0.5, indicating positive parameter drift). Then, the impact propagation process in the network is initiated. The propagation model uses a variant of the linear threshold model: each node has an activation threshold, and its current degree of influence is the weighted sum of the state offsets of all activated upstream nodes pointing to it, multiplied by the corresponding edge weights. If this weighted sum exceeds the node's threshold, the node is activated, and its state offset is calculated using a response function based on its node type and the sum of upstream influences. This process iterates until no new nodes are activated or the preset maximum propagation step count is reached. Each simulation records a propagation path, including the initial node list and the node sequence arranged in activation order. By executing a large number (10,000) of independent simulations, a set of simulated events covering multiple risk sources and propagation scenarios is obtained.

[0102] The iterative formula for the propagation of network disturbances caused by technological influence is shown below:

[0103] ;

[0104] in, For time step ,node State offset. For nodes The preset initial state offsets for the initial disturbance nodes are as follows: for supply shortage nodes, the state offset is 1, indicating complete failure; for quality mutation nodes, the state offset is 0.5, indicating positive parameter drift. The initial set of perturbation nodes is randomly generated in each Monte Carlo simulation. Sigmoid function For time step The set of all activated nodes. For the node Pointing to node The weight of the edge. For nodes State offset for nodes The intensity of the impact, node The activation threshold For other situations.

[0105] Step S52: Based on the set of simulated events and the constraint network, evaluate the degree of violation of the performance constraint boundary by the transmission path. By calculating the degree of violation of the constraint boundary by each node on the transmission path and the cumulative effect of the violation, obtain the constraint violation intensity and the set of key violation nodes for each simulated transmission path.

[0106] Understandably, this step combines the aforementioned set of simulated events with a constraint network injected with market constraints, and performs a process to assess the degree of violation of performance constraint boundaries by the transmission path. For each simulated transmission path, this process checks whether the state change value of each node on the path exceeds the parameter boundary set by the constraint network for that node, which is derived from market tolerance, and calculates the degree to which it exceeds the boundary. Then, this boundary violation effect is accumulated along the transmission path, and finally, a quantitative constraint violation intensity index is calculated for each simulated path, and the set of key nodes that cause major violations is identified, thereby transforming the abstract "impact" into a specific impact on the market-defined product performance red line.

[0107] For each propagation path in the event set, the system traverses every activated node along the path. The state offset obtained from the node simulation is converted into a perturbation of the actual value of the technical parameter represented by the node. Next, the system checks whether the nominal value of the node parameter falls within the feasible region defined for that node by the constraint network. If it does not, the "Euclidean distance" of the boundary violation is calculated. The degree of boundary violation of a single node is defined as the absolute value of exceeding the upper or lower bound.

[0108] The formula for calculating the constraint failure strength of a single conduction path is as follows:

[0109] ;

[0110] in, For the transmission path The overall constraint failure strength, For point The degree to which the parameter value violates its own feasible region is unique. Let be the cascade violation effect coefficient, a non-negative scalar value, taken as 0.3. To indicate in the path In the middle, node It is a node The direct upstream precursor, For the node Pointing to node The weight of the edge. For point The degree to which the parameter value violates its own feasible region is unique. For point The importance weights are defined, and the formula for calculating the severity of a single violation is as follows:

[0111] ;

[0112] in, For nodes The corresponding nominal value of the parameter, and These are the upper and lower bounds of the defined feasible region, respectively.

[0113] Step S53: Based on the constraint failure intensity and key failure node set of each simulated transmission path, search and score the preset historical scheduling schemes to obtain multiple preventive scheduling plans with scores greater than the preset threshold.

[0114] Understandably, this step searches and scores a pre-defined historical scheduling scheme library or strategy space based on the constraint damage intensity and key damage node set of each simulated transmission path. The logic is to evaluate the potential effectiveness of each existing scheduling scheme in blocking high-damage paths or reinforcing key damage nodes, to comprehensively score the performance of each scheme for different risk scenarios, and to select schemes with scores exceeding a pre-defined reliability threshold, forming a set of preventive scheduling plans that can effectively resist identified high-risk transmission paths, thus providing high-quality candidate strategies for the final decision optimization.

[0115] The comprehensive scoring formula for historical scheduling schemes is as follows:

[0116] ;

[0117] in, Historical scheduling scheme Overall score For all original constraint failure strengths Exceeding the high-risk threshold The set of conduction paths, For scheduling scheme For path The relief efficacy score.

[0118] Step S6: Optimize the multi-stage decision sequence coupled with the product release cycle according to the preventive scheduling plan, wherein the scheduling instruction sequence is selected with the goal of minimizing the total switching cost and meeting the network resilience threshold of each stage.

[0119] Understandably, this step places discrete scheduling plans for single risk scenarios within the dynamic context of the entire new product launch cycle for comprehensive optimization. It not only considers the effectiveness of individual plans at the current stage, but more importantly, assesses the restructuring costs and time losses incurred when switching between different plans at different stages. By finding the optimal combination of plans that meets the minimum resilience requirements of each stage and minimizes the total switching cost within a multi-stage sequence, it ultimately outputs a set of executable scheduling instructions strictly bound to key product launch milestones, ensuring the global optimality and practical operability of the strategy. In this step, step S6 includes steps S61 and S62.

[0120] Step S61: Based on the preventive scheduling plan and product release cycle, perform applicability and switching cost assessment. By calculating the degree of achievement of the network resilience level required to maintain this stage, and the reorganization resource and time cost required to switch from the previous stage to this plan, the assessment vector of each preventive scheduling plan is obtained.

[0121] Understandably, this step calculates the extent to which each contingency plan can maintain the network resilience required for the supply chain at a specific stage, and assesses its effectiveness. It also evaluates the resource consumption and time delays caused by operations such as supplier reallocation, inventory reconfiguration, and production plan reorganization when switching from the scheduling plan implemented in the previous stage to the current contingency plan. Thus, an evaluation vector containing two dimensions, "effectiveness" and "switching cost," is calculated for each contingency plan at different stages.

[0122] Specifically, for each contingency plan and each stage, a pre-trained LSTM neural network prediction model is used. Inputting the feature vector of the contingency plan and the environmental characteristics of the stage, the model outputs a predicted level of resilience that the supply chain network can achieve after the implementation of the contingency plan in that stage. This predicted value is divided by the preset minimum resilience threshold for that stage to obtain the effectiveness of the contingency plan in that stage. Simultaneously, each contingency plan is represented as a multi-dimensional resource allocation vector. For any two contingency plans and adjacent stages, the absolute difference in their resource allocation values ​​across each dimension is calculated, multiplied by the unit adjustment cost coefficient of that dimension's resource in that stage, and summed to obtain the switching cost from the previous stage contingency plan to the next stage contingency plan.

[0123] In this step, the formula for calculating the degree of effectiveness is as follows:

[0124] ;

[0125] in, For the contingency plan in time The degree of effectiveness achieved For the contingency plan in time After implementation, the predicted level of supply chain network resilience, Time during the product launch cycle The minimum required network resilience threshold.

[0126] The formula for calculating switching costs is as follows:

[0127] ;

[0128] in, From time contingency plan Switch to time contingency plan Total switching cost required The dimension of the contingency plan vector. For dimension number, In time Adjustment of the Cost coefficient per unit change of resource type Contingency Plan In the Configuration values ​​on dimensional resources, Contingency Plan In the Configuration values ​​on dimensional resources.

[0129] Step S62: Based on the evaluation vector of the preventive scheduling plan, perform multi-stage optimal decision sequence search and construction processing. Take each stage of the product release cycle as the time series, the optional scheduling plan as the state node, and the switching cost of the plan between the previous and next stages as the edge weight to construct a directed acyclic graph. Then, use dynamic programming to search for the plan sequence with the lowest total switching cost under the constraint of the minimum network resilience threshold of each stage to obtain the optimal stage plan sequence.

[0130] This step treats the entire product release cycle as a time series, with each point in time corresponding to a stage. The available contingency plans for each stage are considered state nodes of that stage. The switching cost in the evaluation vector corresponding to switching from one contingency plan in the previous stage to another in the next stage is defined as the weight of the directed edge connecting these two state nodes. This constructs a multi-stage, multi-state directed acyclic graph (DAG). A dynamic programming algorithm is then used to search backwards from the beginning of the graph, finding a coherent contingency plan switching path from the initial stage to the final stage. This path must satisfy the following conditions: the "effectiveness" of the selected contingency plan in each stage is not lower than the preset minimum network resilience threshold for that stage; and the sum of the weights of all edges on the entire path, i.e., the total switching cost, is minimized. Finally, the optimal staged contingency plan sequence represented by this path is output, thus achieving optimization of the overall scheduling switching cost while ensuring the resilience requirements of each stage throughout the entire cycle.

[0131] The dynamic programming state transition equation is shown below:

[0132]

[0133] in, The state value is a dynamic programming term, representing the final selected plan. As time When the contingency plan is in place, the accumulated minimum total switching cost, This is a collection of all contingency plans. Contingency Plan In the stage The degree of effectiveness achieved The state value is a dynamic programming term, representing the final selected plan. As time When the contingency plan is in place, the accumulated minimum total switching cost, From time contingency plan Switch to time contingency plan Total switching cost required Contingency Plan In time The degree of effectiveness achieved.

[0134] Example 2:

[0135] like Figure 2 As shown, this embodiment provides an intelligent supply chain scheduling system that integrates large-scale models. See [link / reference]. Figure 2 The system includes an acquisition unit 701, an extraction unit 702, a construction unit 703, a mapping unit 704, a simulation unit 705, and an optimization unit 706.

[0136] The acquisition unit 701 is used to acquire the timing sensor signal stream that records the component testing process on the assembly and testing production line, the original documents of batch engineering change notices from all component suppliers, and the unstructured text stream data on social media platforms regarding competitor feature experience and fault feedback.

[0137] Extraction unit 702 is used to extract and characterize the implicit features of component batch quality based on the time-series sensor signal stream, and construct an implicit feature vector characterizing each batch of components.

[0138] Construction unit 703 is used to perform adaptive construction processing of change impact network based on the original change notification document and the implicit feature vector of the component, and generate a technology impact network that describes the transmission relationship of change in the manufacturing chain with components and process parameters as nodes and offsets as edge weights.

[0139] The mapping unit 704 is used to perform reverse mapping processing of market demand disturbances and technical constraints based on the unstructured text stream data and the technology influence network, and to back-couple the market demand disturbances after reverse mapping as new constraints to the technology influence network to obtain a constraint network that reshapes the boundaries of network nodes.

[0140] The simulation unit 705 is used to simulate supply chain disturbances based on the technology's impact on the network, calculate the degree of damage to the boundary conditions of nodes in the constrained network caused by the simulated transmission path, and generate a preventive scheduling plan.

[0141] The optimization unit 706 is used to perform multi-stage decision sequence optimization coupled with the product release cycle according to the preventive scheduling plan, wherein the scheduling instruction sequence is selected with the goal of minimizing the total switching cost and meeting the network resilience threshold of each stage.

[0142] It should be noted that the specific methods by which each module performs operations in the system described in the above embodiments have been described in detail in the embodiments related to the method, and will not be elaborated here.

[0143] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0144] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A method for intelligent supply chain scheduling that integrates large-scale models, characterized in that, include: Acquire timing sensor signal streams recording component testing processes on assembly and testing production lines, original documents of batch engineering change notifications from all component suppliers, and unstructured text streams of competitor feature experience and fault feedback from social media platforms; Based on the time-series sensor signal stream, implicit feature extraction and characterization processing of component batch quality are performed to construct implicit feature vectors representing components in each batch. Based on the original change notification document and the implicit feature vector of the component, an adaptive construction process of the change impact network is performed. The technology impact network that describes the transmission relationship of the change in the manufacturing chain is generated with the component and process parameters as nodes and the offset as the edge weight. The unstructured text stream data and the technology influence network are subjected to inverse mapping of market demand disturbances and technology constraints. The market demand disturbances after inverse mapping are then coupled back to the technology influence network as new constraints, resulting in a constraint network that reshapes the boundaries of network nodes. Based on the aforementioned technology, supply chain disturbance simulations are performed on the network, and the degree of disruption to the boundary conditions of nodes in the constrained network caused by the simulated transmission paths is calculated to generate preventive scheduling plans. Based on the aforementioned preventive scheduling plan, a multi-stage decision sequence optimization coupled with the product release cycle is performed, wherein the scheduling instruction sequence is selected with the goal of minimizing the total switching cost and meeting the network resilience threshold of each stage.

2. The intelligent supply chain scheduling method based on a fusion large model according to claim 1, characterized in that, Based on the time-series sensor signal stream, implicit feature extraction and characterization of component batch quality are performed, including: Based on the time-series sensor signal stream, multi-process signal coupling and decomposition processing is performed. By aligning and decomposing the sensor signals of different test processes in the time-frequency domain, the mixed signal components characterizing the inherent properties of the components, the contact state of the fixture, and the environmental noise are separated to obtain the decoupled component body response signal sequence. Based on the response signal sequence of the component body, cross-batch abnormal pattern clustering and tracing processing are performed, and the processed abnormal patterns are associated and mapped with the supplier's production batch identifier to obtain an abnormal pattern feature cluster with supplier batch tracing label. Based on the abnormal pattern feature cluster with supplier batch traceability tags, a process robustness implicit coding vector is synthesized to obtain an implicit feature vector characterizing the overall quality of the components in that batch.

3. The intelligent supply chain scheduling method based on a fusion large model according to claim 1, characterized in that, An adaptive construction process for the change impact network is performed based on the original change notification document and the implicit feature vectors of the components, using components and process parameters as nodes and offsets as edge weights, including: Based on the original document of the change notification and the implicit feature vector, the parameters associated with the change items and the affected batches are extracted and processed. The extracted associated parameters are then mapped to the corresponding implicit feature vectors of the components to obtain the set of change parameters and their associated batches and feature vectors. Based on the changed parameter set and its associated batches and feature vectors, the feature offset caused by the change is calculated. The changed parameter values ​​are input into the feature vector generation model to calculate the direction and magnitude of the disturbance to the original associated implicit feature vectors, and a quantified feature offset vector is obtained. Based on the quantized feature offset vector, a directed weighted network is constructed. Using component batches and manufacturing process parameters as nodes, and the disturbance direction and magnitude represented by the feature offset vector as the weight and direction of the edges, a technology influence network for the multi-level transmission relationship of change notifications in the manufacturing chain is constructed.

4. The intelligent supply chain scheduling method based on a fusion large model according to claim 1, characterized in that, The unstructured text stream data and the technology influence network are subjected to inverse mapping of market demand disturbances and technological constraints. The inversely mapped market demand disturbances are then back-coupled into the technology influence network as new constraints, including: Based on the unstructured text stream data, market sentiment quantification processing is performed on the functional defects of competitors. By performing fine-grained sentiment and entity analysis on the text stream, the intensity of collective dissatisfaction expressed by users regarding the performance defects of competitors is identified and quantified, resulting in a market sentiment intensity vector. Based on the market sentiment intensity vector and the technology influence network, the sentiment vector is mapped to the network technology nodes and constrained. By converting the market sentiment intensity vector into a tolerance threshold for the fluctuation range of relevant technology node parameters, a set of performance constraints for the corresponding network nodes is obtained. Based on the set of performance constraints, the constraints are reverse-injected into the technology influence network and the boundary is reshaped. By applying the tolerance threshold in the set of performance constraints as a new boundary condition to the corresponding parameter node in the technology influence network, the feasible domain of the node state is modified, and the constraint network after the node boundary is reshaped is obtained.

5. The intelligent supply chain scheduling method based on a fusion large model according to claim 1, characterized in that, Based on the aforementioned technology, supply chain disturbance simulations are performed on the network, and the degree of disruption to the boundary conditions of nodes in the constrained network caused by the simulated transmission paths is calculated. Preventive scheduling plans are then generated, including: Based on the technology, the network is used to simulate disturbance events of typical supply chain risks. In this process, node state anomalies representing key component supply shortages, batch quality mutations, or emergency engineering changes are triggered with a preset probability. The process of the abnormal state spreading along the transmission relationship defined by the network edge weights is simulated to obtain a set of simulated events including the initial disturbance node and subsequent transmission paths. The impact of the transmission path on the performance constraint boundary is evaluated based on the set of simulated events and the constraint network. By calculating the degree of violation of the constraint boundary by each node on the transmission path and the cumulative effect of the violation, the constraint violation intensity and the set of key violation nodes of each simulated transmission path are obtained. Based on the constraint failure intensity and key failure node set of each simulated transmission path, the preset historical scheduling schemes are searched and scored to obtain multiple preventive scheduling plans with scores greater than the preset threshold.

6. An intelligent supply chain scheduling system integrating a large model, characterized in that, include: The acquisition unit is used to acquire timing sensor signal streams that record component testing processes on the assembly and testing production line, original documents of batch engineering change notices from all component suppliers, and unstructured text stream data from social media platforms regarding competitor feature experience and fault feedback. The extraction unit is used to extract and characterize the implicit features of component batch quality based on the time-series sensor signal stream, and to construct an implicit feature vector characterizing each batch of components. The construction unit is used to adaptively construct the change impact network based on the original change notification document and the implicit feature vector of the component, and generate a technology impact network that describes the transmission relationship of the change in the manufacturing chain, with the component and process parameters as nodes and the offset as the edge weight. The mapping unit is used to perform reverse mapping processing of market demand disturbances and technical constraints based on the unstructured text stream data and the technology influence network, and to back-couple the reverse-mapped market demand disturbances as new constraints to the technology influence network to obtain a constraint network that reshapes the boundaries of network nodes. The simulation unit is used to simulate supply chain disturbances based on the technology's impact on the network, calculate the degree of damage to the boundary conditions of nodes in the constrained network caused by the simulated transmission path, and generate preventive scheduling plans. The optimization unit is used to perform multi-stage decision sequence optimization coupled with the product release cycle according to the preventive scheduling plan, wherein the scheduling instruction sequence is selected with the goal of minimizing the total switching cost and meeting the network resilience threshold of each stage.

7. The intelligent supply chain scheduling system based on the fusion of large models according to claim 6, characterized in that, The extraction unit includes: The first extraction subunit is used to perform multi-process signal coupling and decomposition processing based on the time-series sensor signal stream. By aligning and decomposing the sensor signals of different test processes in the time-frequency domain, the mixed signal components characterizing the inherent properties of the components, the contact state of the fixture, and the environmental noise are separated to obtain a decoupled component body response signal sequence. The second extraction subunit is used to perform cross-batch abnormal pattern clustering and tracing processing based on the component body response signal sequence, and to associate and map the processed abnormal patterns with the supplier production batch identifier to obtain an abnormal pattern feature cluster with supplier batch tracing label. The third extraction subunit is used to synthesize the implicit encoding vector of process robustness based on the abnormal pattern feature cluster with supplier batch traceability label, so as to obtain the implicit feature vector characterizing the overall quality of the batch of components.

8. The intelligent supply chain scheduling system based on the fusion of large models according to claim 6, characterized in that, The building unit includes: The first construction subunit is used to extract the parameters associated with the change items and the affected batches based on the original document of the change notification and the implicit feature vector, and to map the extracted associated parameters to the corresponding implicit feature vectors of the components to obtain the set of change parameters and their associated batches and feature vectors. The second construction subunit is used to calculate the feature offset caused by the change based on the changed parameter set and its associated batch and feature vector. By simulating the changed parameter values ​​and inputting them into the feature vector generation model, the perturbation direction and magnitude of the original associated implicit feature vector are calculated to obtain the quantified feature offset vector. The third construction subunit is used to construct a directed weighted network based on the quantized feature offset vector. Using component batches and manufacturing process parameters as nodes, and the disturbance direction and magnitude represented by the feature offset vector as the weight and direction of the edges, a technology influence network for the multi-level transmission relationship of change notifications in the manufacturing chain is constructed.

9. The intelligent supply chain scheduling system based on the fusion of large models according to claim 6, characterized in that, The mapping unit includes: The first mapping subunit is used to perform market sentiment quantification processing on the unstructured text stream data in response to the functional defects of competitors. By performing fine-grained sentiment and entity analysis on the text stream, it identifies and quantifies the intensity of collective dissatisfaction expressed by users regarding the performance defects of competitors, and obtains a market sentiment intensity vector. The second mapping subunit is used to perform mapping and constraint transformation of the sentiment vector to network technology nodes based on the market sentiment intensity vector and the technology influence network. By converting the market sentiment intensity vector into a tolerance threshold for the fluctuation range of relevant technology node parameters, a set of performance constraints for the corresponding network node is obtained. The third mapping subunit is used to perform reverse injection and boundary reshaping of the constraint conditions on the technology influence network according to the performance constraint condition set. By applying the tolerance threshold in the performance constraint condition set as a new boundary condition to the corresponding parameter node in the technology influence network, the feasible domain of the node state is modified, and the constraint network after the node boundary is reshaped is obtained.

10. The intelligent supply chain scheduling system based on the fusion of large models according to claim 6, characterized in that, The simulation unit includes: The first simulation subunit is used to simulate and process disturbance events of typical supply chain risks based on the technology-affected network. In this subunit, node state anomalies representing key component supply shortages, batch quality mutations, or emergency engineering changes are triggered with a preset probability. The simulation process of the abnormal state spreading along the transmission relationship defined by the network edge weights is used to obtain a set of simulated events containing the initial disturbance node and subsequent transmission paths. The second simulation subunit is used to evaluate the degree of violation of the performance constraint boundary by the transmission path based on the set of simulated events and the constraint network. By calculating the degree of violation of the constraint boundary by each node on the transmission path and the cumulative effect of the violation, the constraint violation intensity and the set of key violation nodes of each simulated transmission path are obtained. The third simulation subunit is used to search and score the preset historical scheduling schemes based on the constraint failure intensity and the set of key failure nodes of each simulated transmission path, and obtain multiple preventive scheduling plans with scores greater than the preset threshold.

Images