An AI vertical model-based industrial supply chain intelligent regulation method, medium and system

Abstract

The application provides an industrial supply chain intelligent regulation and control method based on an AI vertical field model, a medium and a system, and belongs to the technical field of artificial intelligence. The application stores the cleaned data into a time series database of a log-structured merge tree, extracts historical supply interruption data, supplier capacity fluctuation data and transportation delay data, inputs the data into an elastic early warning model based on a Transformer architecture to calculate a supply interruption probability value within 72 hours, generates a production scheduling plan and a transportation scheduling scheme by using a supply chain balance scheduling algorithm constrained by thermodynamic entropy increase after receiving a warning signal, calculates the total carbon emission of the whole link by using a graph neural network and outputs an emission reduction optimization scheme, and feeds a supplementary feature vector to the model for parameter updating according to an actual delivery delay rate deviation, so that the technical problem of resource scheduling response lag caused by insufficient supply chain interruption prediction accuracy is solved.

Description

Technical Field

[0001] This invention belongs to the field of artificial intelligence technology, and specifically relates to an intelligent control method, medium, and system for industrial supply chain based on an AI vertical domain model. Background Technology

[0002] In the field of industrial supply chain management, traditional methods use linear regression models or time series analysis based on historical statistical data to predict supply disruption risks. These methods trigger manual scheduling decisions by setting fixed safety stock thresholds, relying on expert experience to formulate production plans and logistics routes. However, in the current multi-source heterogeneous data environment, OPC UA protocol device data and MQTT protocol IoT data use different message structures and semantic representations, making it impossible to directly use the collected data for unified analysis. This necessitates the manual writing of numerous data conversion scripts. Traditional prediction models only consider the historical frequency of supply disruptions at a single time scale, ignoring the complex dependencies between multiple dimensions such as supplier capacity fluctuations, transportation delays, and weather factors. This leads to a significant decrease in prediction accuracy when facing sudden supply chain disruptions, resulting in delayed or false alarms in early warning signals. In other words, existing technologies suffer from insufficient accuracy in supply chain disruption prediction, leading to delayed resource scheduling responses. Summary of the Invention

[0003] In view of this, the present invention provides an intelligent control method, medium and system for industrial supply chain based on AI vertical domain model, which can solve the technical problem of insufficient accuracy in supply chain interruption prediction leading to delayed resource scheduling response in the prior art.

[0004] The present invention is implemented as follows: The first aspect of the present invention provides an intelligent control method for industrial supply chains based on an AI vertical domain model. This method deploys multi-protocol data acquisition terminals to collect equipment operating parameters, material inventory levels, and carbon emissions. These data are then converted into standard format data using a protocol parsing engine. A supply chain knowledge graph is constructed to establish entity relationships. The cleaned data is written into a time-series database. Historical supply disruption data, supplier capacity fluctuation data, and transportation delay data are extracted as feature vectors and input into an elastic early warning model to calculate the probability of supply disruption and output an early warning signal. Upon receiving the early warning signal, a supply chain balance scheduling algorithm based on thermodynamic entropy increase constraints is used to calculate resource allocation schemes and output production and transportation scheduling schemes. Based on the production schedule, a graph neural network is used to calculate the total carbon emissions across the entire supply chain and output emission reduction optimization schemes. Based on the deviation between the actual delivery delay rate and the predicted delivery delay rate, supplementary feature vectors are fed back to the elastic early warning model for parameter updates.

[0005] The protocol parsing engine extracts semantic information from data fields by parsing the message structure of OPC UA protocol data and MQTT protocol data, and converts them into standard format data containing timestamps, device identifiers, parameter types, and values ​​according to predefined mapping rules.

[0006] The standard format data is stored using JSON object representation.

[0007] The steps involved in constructing a supply chain knowledge graph specifically involve using a two-way attention disambiguation model to identify entities and extract relationships from text information in standard format data, and establishing the association and lineage relationships between material identifiers, supplier identifiers, production work order identifiers, and transportation order identifiers.

[0008] The bidirectional attention disambiguation model receives word vector sequences from industrial texts in its input layer, extracts contextual semantic features in its bidirectional long short-term memory network layer, calculates the similarity weight between each word and the candidate entity in its attention layer, and combines the semantic vectors with the prior probabilities of the knowledge base through Bayes' theorem to obtain the disambiguation result in its output layer.

[0009] The training dataset for the bidirectional attention disambiguation model extracts samples containing entity and relation annotations from historical work order texts. For unannotated samples, a remote supervised learning method is used to automatically generate pseudo-labels. The annotated samples and pseudo-labeled samples are then merged to form the training dataset.

[0010] The cleaning process specifically involves eliminating outliers in equipment operating parameters using the three-times-standard-deviation principle and desensitizing supplier identifiers.

[0011] The time-series database adopts a log structure merge tree architecture, including a memory buffer table and a disk storage table. Newly collected standard format data is first written to the memory buffer table, and when the memory buffer table reaches a preset capacity threshold, it is batch-written to the disk storage table.

[0012] The disk storage table uses a columnar storage format, performs run-length encoding compression on duplicate data, and the query operation uses a Bloom filter to determine whether standard format data exists before actually reading it.

[0013] The resilient early warning model adopts the Transformer architecture. The input layer adds the feature vector and the position encoding and then inputs it into the multi-head attention layer. The multi-head attention layer calculates the dependencies at different time scales in parallel. The output layer maps the hidden state to the supply interruption probability value through a fully connected neural network.

[0014] The preset threshold for the probability of supply disruption is set as the probability value corresponding to the 95th percentile, based on the statistical distribution of historical supply disruption events.

[0015] The thermodynamic entropy increase constraint-based supply chain balancing scheduling algorithm maps the inventory levels of each node in the supply chain to the particle distribution state of a thermodynamic system. The information entropy of the inventory distribution is defined as the negative value of the logarithmic weighted sum of the proportions of each node's inventory to the total inventory. The thermodynamic entropy increase constraint-based supply chain balancing scheduling algorithm introduces a temperature parameter to control the probability of accepting a suboptimal solution. The initial temperature is set to a relatively high value, and after each iteration, the temperature decreases according to an exponential decay function.

[0016] The graph neural network includes graph convolutional layers and pooling layers. The graph convolutional layers aggregate the feature vectors of raw material batch nodes, transportation path nodes, and production process nodes to the central node. The pooling layers extract the carbon emission contribution of key nodes. The output layer sums the carbon emissions of all nodes to obtain the total carbon emissions of the entire chain.

[0017] Among them, the emission reduction optimization plan includes instructions to replace high-carbon transportation routes with low-carbon transportation routes.

[0018] Specifically, the parameter update step involves concatenating the supplementary feature vector composed of weather factor data and traffic congestion data with the feature vector to form an extended feature vector, retraining the input layer weight matrix of the elastic early warning model, keeping the parameters of other network layers unchanged, and minimizing the mean square error between the prediction result and the actual result through the gradient descent algorithm.

[0019] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the above-described intelligent control method for an industrial supply chain based on an AI vertical domain model.

[0020] A third aspect of the present invention provides an intelligent control system for industrial supply chain based on an AI vertical domain model, comprising the aforementioned computer-readable storage medium, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor for executing program instructions stored in the computer-readable storage medium.

[0021] This invention employs a resilient early warning model to replace traditional statistical forecasting methods. Through a Transformer architecture with multi-head attention layers, it parallelly computes the dependencies between historical supply disruption data, supplier capacity fluctuation data, and transportation delay data at different time scales. This extracts nonlinear correlation patterns between multi-dimensional feature vectors, extending the prediction window for supply disruption probability from 24 hours to 72 hours, significantly improving the lead time of the warning signal. The resilient early warning model prevents the vanishing gradient problem during deep network training through residual connections and layer normalization mechanisms, ensuring the model can learn feature dependencies over long time spans. When the deviation between the actual delivery delay rate and the predicted value exceeds a threshold, the system automatically feeds back weather and traffic congestion data as supplementary feature vectors to the model for parameter updates, achieving continuous optimization of prediction accuracy. In summary, this invention solves the technical problem mentioned in the background art where insufficient accuracy in supply chain disruption prediction leads to delayed resource scheduling response. Attached Figure Description

[0022] Figure 1 This is a flowchart of the method of the present invention.

[0023] Figure 2 This is a time-series distribution diagram of data collection over 72 hours.

[0024] Figure 3 This is a network structure diagram of the supply chain knowledge graph.

[0025] Figure 4 This is a curve predicting the probability of supply disruption.

[0026] Figure 5 A comparison chart showing the inventory distribution before and after optimization.

[0027] Figure 6 This is a diagram showing the percentage of carbon emissions across the entire value chain. Detailed Implementation

[0028] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below.

[0029] like Figure 1 The diagram shown is a flowchart of an intelligent control method for industrial supply chains based on an AI vertical domain model, provided by the first aspect of this invention. This method includes the following steps:

[0030] S01. Deploy multi-protocol data acquisition terminals in supplier workshops and logistics nodes. Use a protocol parsing engine to convert OPC UA protocol data and MQTT protocol data into standard format data to collect equipment operating parameters, material inventory and carbon emissions.

[0031] S02. Construct a supply chain knowledge graph, use a two-way attention disambiguation model to identify entities and extract relationships from text information in standard format data, and establish the association lineage relationship between material identifiers, supplier identifiers, production work order identifiers and transportation order identifiers.

[0032] S03. Clean the standard format data, remove outliers in the equipment operating parameters by using the three-times-standard-deviation principle, desensitize the supplier identifier, and write the processed data into a time-series database based on a log structure merging tree.

[0033] S04. Extract historical supply interruption data, supplier capacity fluctuation data and transportation delay data from the time series database as feature vectors, input them into the elastic early warning model, calculate the supply interruption probability value within 72 hours, and output an early warning signal when the supply interruption probability value exceeds the preset threshold.

[0034] S05. After receiving the early warning signal, extract sales order data and real-time inventory data, calculate resource allocation scheme based on the supply chain balance scheduling algorithm with thermodynamic entropy increase constraint, and output production plan and transportation scheduling scheme as control results;

[0035] S06. Collect raw material batch data, transportation route data and production process parameters according to the production schedule, calculate the total carbon emissions of the whole chain through graph neural network correlation analysis, and output the emission reduction optimization scheme when the total carbon emissions of the whole chain exceed the carbon emission threshold.

[0036] S07. After executing the production schedule and transportation scheduling plan, collect the actual delivery delay rate, calculate the deviation between the actual delivery delay rate and the predicted delivery delay rate, and when the deviation exceeds 5%, feed back weather factor data and traffic congestion data as supplementary feature vectors to the elastic early warning model for parameter updates.

[0037] The protocol parsing engine extracts semantic information from data fields by parsing the message structure of OPC UA protocol data and MQTT protocol data. According to predefined mapping rules, it converts OPC UA protocol data and MQTT protocol data into standard format data containing timestamps, device identifiers, parameter types, and values. The standard format data is stored using JSON object representation.

[0038] The bidirectional attention disambiguation model is structured as follows: the input layer receives a sequence of word vectors from industrial text; a bidirectional long short-term memory network layer extracts contextual semantic features; the attention layer calculates the similarity weight between each word and candidate entities in the industrial dictionary; and the output layer combines semantic vectors and prior probabilities from the knowledge base using a Bayesian formula to obtain the disambiguation result. The steps for establishing the training dataset for the bidirectional attention disambiguation model include extracting samples containing entity and relation annotations from historical work order texts; automatically generating pseudo-labels for unlabeled samples using a remote supervised learning method; and merging labeled and pseudo-labeled samples to form the training dataset. The training steps for the bidirectional attention disambiguation model include dividing the training dataset into a training set and a validation set in an 8:2 ratio; using a cross-entropy loss function to measure the difference between the predicted results and the actual annotations; updating network parameters using a backpropagation algorithm; and stopping training when the F1 value of the validation set does not improve for five consecutive cycles.

[0039] The time-series database architecture of the log structure merging tree includes a memory buffer table and a disk storage table. Newly collected standard format data is first written to the memory buffer table. When the memory buffer table reaches a preset capacity threshold, it is batch-written to the disk storage table. The disk storage table adopts a columnar storage format, and duplicate data is compressed using run-length encoding. The query operation uses a Bloom filter to determine whether the standard format data exists before actually reading it.

[0040] The resilient early warning model uses a Transformer architecture instead of a long short-term memory network. The input layer adds a feature vector composed of historical supply disruption data, supplier capacity fluctuation data, and transportation delay data to the location encoding and then inputs it into a multi-head attention layer. The multi-head attention layer calculates the dependencies at different time scales in parallel and prevents gradient vanishing through residual connections and layer normalization. The output layer maps the hidden state to a supply disruption probability value through a fully connected neural network. The preset threshold is set to the probability value corresponding to the 95th percentile based on the statistical distribution of historical supply disruption events. The predicted delivery delay rate is calculated by the resilient early warning model based on the feature vector.

[0041] The thermodynamic entropy increase constraint-based supply chain balancing scheduling algorithm maps the inventory levels of each node in the supply chain to the particle distribution state of a thermodynamic system. The information entropy of the inventory distribution is defined as the negative value of the logarithmic weighted sum of the proportions of each node's inventory to the total inventory. A smaller information entropy indicates a more balanced inventory distribution. The algorithm introduces a temperature parameter to control the probability of accepting suboptimal solutions. The initial temperature is set to a relatively high value, and after each iteration, the temperature decreases according to an exponential decay function. During the high-temperature phase, the system, based on the Boltzmann distribution, accepts resource allocation schemes that increase information entropy with a higher probability, achieving a wide-range search and avoiding getting trapped in local optima. During the low-temperature phase, the system accepts suboptimal solutions with a lower probability, achieving a refined search for the optimal solution. The mapping relationship between supply chain systems and thermodynamic systems lies in the fact that the uneven distribution of inventory corresponds to the non-equilibrium state of thermodynamic systems. The process of the system spontaneously tending towards entropy increase, i.e., the balanced distribution of inventory, is analogous to the spontaneous entropy increase process described by the second law of thermodynamics. The simulated annealing process controlled by temperature parameters allows the system to gradually evolve from a high-entropy non-equilibrium state to a low-entropy equilibrium state, which corresponds to the supply chain adjusting from a state of concentrated inventory risk to a state of balanced configuration. The supply chain balancing scheduling algorithm constrained by thermodynamic entropy increases reduces the factorial computational complexity of traditional traversal search to the level of square multiplied by logarithm.

[0042] The graph neural network includes graph convolutional layers and pooling layers. The graph convolutional layers aggregate the feature vectors of raw material batch nodes, transportation path nodes, and production process nodes to a central node. The pooling layers extract the carbon emission contribution of key nodes. The output layer sums the carbon emissions of all nodes to obtain the total carbon emissions of the entire chain. The carbon emission threshold is set according to industry standards. The emission reduction optimization scheme includes instructions to replace high-carbon transportation paths with low-carbon transportation paths.

[0043] The parameter update step includes concatenating a supplementary feature vector composed of weather factor data and traffic congestion data with a feature vector composed of historical supply disruption data, supplier capacity fluctuation data and transportation delay data to form an extended feature vector; retraining the input layer weight matrix of the elastic early warning model; keeping the other network layer parameters of the elastic early warning model unchanged; minimizing the mean square error between the prediction result and the actual result of the extended feature vector using the gradient descent algorithm; and completing the parameter update when the mean square error is reduced to below a preset convergence threshold.

[0044] The production schedule includes the start-up time of production equipment, the processing sequence of production batches and the release time of materials. The transportation scheduling plan includes the departure time of transport vehicles, the selection of transportation routes and the allocation of target warehouses. The production schedule and transportation scheduling plan are output as control results to the production execution system and the logistics management system.

[0045] The specific implementation methods of the above steps are described in detail below.

[0046] The specific implementation of step S01 involves deploying a multi-protocol data acquisition terminal at the production equipment end of the supplier's workshop. This terminal first connects to a programmable logic controller supporting the OPC UA protocol via an Ethernet interface to acquire equipment operating parameters such as injection molding machine temperature parameters and stamping press pressure parameters. Simultaneously, it accesses an IoT sensor supporting the MQTT protocol via a wireless network to collect data on raw material inventory and production process parameters. Emissions. Upon receiving the message, the protocol parsing engine parses the node identifier and data type of OPC UA protocol data according to the IEC 62541 standard, extracting semantic information such as device number, measurement value, and quality code. For MQTT protocol data, it parses the numerical fields in the message payload according to the topic hierarchy structure. Based on a pre-configured mapping rule table, the parsing engine uniformly converts the heterogeneous data from both protocols into standard format data containing four fields: timestamp, device identifier, parameter type, and value. This standard format data is stored using JSON object representation, facilitating unified parsing by subsequent processing modules. This step aims to eliminate data format differences caused by multiple communication protocols in industrial settings, achieving standardized access to heterogeneous data sources.

[0047] The specific implementation of step S02 involves constructing a supply chain knowledge graph to establish semantic relationships between entities. First, fields containing textual information such as material names, supplier names, and work order numbers are extracted from standard format data. This text is then input into a bidirectional attention disambiguation model for processing. The input layer of this model receives a sequence of word vectors transformed using word embedding technology, with each word represented as a 300-dimensional dense vector. A bidirectional long short-term memory network layer scans the word vector sequence in both forward and backward directions, capturing contextual semantic features. The forward network transmits information from the beginning to the end of the sentence, while the backward network transmits information from the end to the beginning. The hidden state vectors from both directions are concatenated to form a complete contextual representation. The attention layer calculates the semantic similarity between the word to be disambiguated and all candidate entities in the industrial domain dictionary. A dot product operation is used to measure the distance between the word vector and the entity vector, and the similarity weight distribution is obtained through softmax normalization. The output layer fuses the semantic vector obtained from the attention layer with the prior probabilities of candidate entities in the knowledge base, calculates the posterior probability using Bayes' theorem, and selects the entity with the highest posterior probability as the disambiguation result. During the model training phase, labeled entity and relationship samples are extracted from historical work order texts as supervised data. For unlabeled samples, a remote supervised learning method is used, automatically generating pseudo-labels by matching known entity pairs in the knowledge base. Manually labeled samples and pseudo-labeled samples are mixed at a 6:4 ratio to form the training dataset. During training, the dataset is divided into training and validation sets at an 8:2 ratio. The cross-entropy loss function is used to measure the difference between the model's predicted entity category distribution and the actual labels. Network parameters are updated using the backpropagation algorithm and the Adam optimizer, with a learning rate set to 0.001. Training is terminated when the F1 score on the validation set fails to improve for five consecutive training epochs. After disambiguation, entities are extracted, and semantic relationships between entities are identified through a relationship extraction module. Finally, a relationship graph is constructed between material identifiers, supplier identifiers, production work order identifiers, and transportation order identifiers. This step establishes structured relationships between various elements of the supply chain, providing semantic support for subsequent relationship analysis and traceability queries.

[0048] The specific implementation of step S03 involves quality cleaning and privacy protection processing of standard format data. For equipment operating parameter data, the mean and standard deviation of all historical data are first calculated. Values ​​deviating from the mean by more than three times the standard deviation are identified as outliers and removed. For example, the historical mean of injection molding machine temperature parameters is 185℃, and the standard deviation is 12℃. Records with temperatures below 149℃ or above 221℃ are considered outliers. For the supplier identification field, a hash function is used to map the real supplier name to a fixed-length hash value, ensuring that the supplier's true identity information is not leaked during data analysis. The hash algorithm used is SHA-256 to ensure the one-way mapping and collision resistance. After cleaning and desensitization, the processed data is written to a time-series database based on a log structure merging tree. This database architecture includes a memory buffer table and multiple layers of disk storage tables. Newly collected data is first appended to the memory buffer table to achieve high throughput performance. When the data volume in the memory buffer table reaches a capacity threshold of 64MB, a flush operation is triggered to batch dump the data to the first-level storage table on the disk. The disk storage table uses a columnar storage format, storing data from the same column contiguously to improve compression ratio and query efficiency. Repeated values ​​are compressed using run-length encoding; for example, if 100 consecutive records show the device status as "running," the encoding would be "running × 100." During a query, a Bloom filter is first used to quickly determine if the timestamp being queried might exist in a particular storage file. The Bloom filter uses multiple hash functions to map the timestamp to a bit array. If all corresponding bits in the bit array are 1, the file may exist and requires further querying; if any bit is 0, the file is skipped, reducing disk I / O overhead. This step ensures data quality, protects business privacy, and provides efficient time-series data storage and retrieval capabilities.

[0049] The specific implementation of step S04 involves constructing a resilient early warning model to predict supply chain disruption risks. First, historical supply disruption event records from the past six months, daily supplier capacity fluctuation data, and transportation delay duration data are extracted from a time-series database. This data is organized into time-series feature vectors, with each time step containing 15 feature dimensions. The resilient early warning model is constructed using a Transformer architecture. The input layer adds the feature vectors to the positional encoding and then inputs them into a multi-head attention layer. The positional encoding is generated using sine and cosine functions, enabling the model to perceive the positional information at different time steps in the time series. The multi-head attention layer has eight attention heads, each independently calculating the attention weights between the query vector, key vector, and value vector. Parallel computation captures the dependencies between different time scales, such as short-term fluctuations and long-term trends. The outputs of the multi-head attention layer are concatenated and further feature-extracted through a feedforward neural network. The model introduces residual connections and layer normalization mechanisms. Residual connections directly add the input of the attention layer to the output, preventing the gradient vanishing problem during deep network training. Layer normalization standardizes the features of each sample to accelerate convergence. The output layer employs a two-layer fully connected neural network. The first layer has 128 neurons with the ReLU activation function, and the second layer has 1 neuron with the Sigmoid activation function, mapping the hidden states to supply disruption probability values ​​between 0 and 1. The preset threshold is determined by analyzing the probability distribution of historical supply disruption events. The 95th percentile of the probability values ​​output by the model during supply disruptions over the past year is calculated and used as the warning threshold, typically between 0.72 and 0.85. When the model calculates that the probability of supply disruption within the next 72 hours exceeds this threshold, the system automatically outputs a warning signal to notify dispatchers to intervene. The model also outputs a predicted delivery delay rate, calculated through regression analysis based on historical transportation delay data in the feature vector. This step aims to identify vulnerable links in the supply chain in advance, providing a window of opportunity for proactive adjustments.

[0050] The specific implementation of step S05 involves initiating a supply chain balancing scheduling algorithm to calculate resource allocation schemes after receiving an early warning signal. First, the demand, delivery deadlines, and real-time inventory data of all currently unfulfilled sales orders, as well as the inventory levels of each warehouse node, are extracted from the business system. The supply chain balancing scheduling algorithm based on thermodynamic entropy increase constraints analogizes the inventory level of each warehouse node in the supply chain network to the distribution of particles at different energy levels in a thermodynamic system. The information entropy of the inventory distribution is defined as the negative value of the logarithmic weighted sum of the proportions of each node's inventory to the total inventory. A lower information entropy indicates an unbalanced distribution when inventory is concentrated in a few nodes, while a higher information entropy indicates a balanced distribution when inventory is evenly distributed across nodes. The algorithm introduces a temperature parameter to control the probability of accepting inferior solutions during optimization. The initial temperature is set to 1000 Kelvin, and after each iteration, the temperature decreases according to an exponential decay function with a decay coefficient of 0.95. During the high-temperature phase, the system, based on the Boltzmann distribution, accepts resource allocation schemes that increase information entropy with a higher probability, even if the scheme temporarily increases transportation costs, avoiding getting trapped in local optima through a large-scale random search. The calculation of state transition probability considers the information entropy difference between two adjacent resource allocation schemes. This difference is divided by the reference information entropy value and then by the ratio of the temperature parameter to the reference temperature value. The negative value of the exponential function is taken to obtain the acceptance probability, where the normalization coefficient of the Boltzmann constant is set to 1. As the temperature gradually decreases and enters the low-temperature stage, the system accepts inferior solutions with a lower probability. The optimization process shifts from global exploration to local fine-grained search, eventually converging to the optimal allocation scheme with lower information entropy that meets order requirements. This algorithm reduces the factorial computational complexity of traditional traversal search methods to the logarithmic level by leveraging the mapping relationship between the thermodynamic system and the supply chain system, enabling the scheduling calculation of large-scale supply chain networks to be completed within an acceptable timeframe. The algorithm outputs a production scheduling plan, including the start-up time of each production device, the processing sequence of each production batch, and the raw material delivery time. It also outputs a transportation scheduling plan, including the departure time of transport vehicles, the selection of driving routes, and the allocation results of target warehouses. The purpose of this step is to quickly generate a control plan that balances inventory balance and cost optimization when a supply risk warning is triggered.

[0051] The specific implementation of step S06 involves collecting and optimizing end-to-end carbon emission data based on the production schedule. According to the production batches determined by the schedule, the source of raw materials used in each batch, the route of the transport vehicles, and the parameters set for the production process are collected. This data is input into a graph neural network as node features for correlation analysis. The graph neural network models raw material batches, transport routes, and production processes as nodes in a graph structure, with edges between nodes representing material flow relationships or process sequences. The graph convolutional layer aggregates the feature vectors of adjacent nodes to the central node through a message passing mechanism. Each node's feature vector contains information such as the carbon emission coefficient and duration of that stage. The graph convolution operation models the cumulative effect of carbon emission propagation in the supply chain network. The pooling layer uses attention pooling to calculate the contribution weight of each node to the total end-to-end carbon emission, extracting the key nodes with the highest carbon emission contribution for subsequent optimization. The output layer sums the carbon emissions of all nodes to obtain the total end-to-end carbon emission from raw material procurement to product delivery, with the unit being kilograms. Equivalent. The carbon emission threshold is set according to industry standards or corporate targets, and is generally taken as 500 kg per unit of product. Equivalent. When the total carbon emissions across the entire supply chain exceed a threshold, the system outputs an emission reduction optimization plan. This plan analyzes the carbon emission contribution of key nodes, identifies the transportation route with the highest carbon emission intensity, and generates execution instructions to replace it with low-carbon transportation routes such as rail transport or new energy vehicle transport. The purpose of this step is to achieve refined accounting and dynamic optimization of the supply chain's carbon footprint.

[0052] The specific implementation of step S07 involves evaluating the effectiveness of the control plan and adaptively updating the model after execution. After the production schedule and transportation scheduling plan are completed, the actual order delivery delay rate is collected from the logistics management system. This delay rate is defined as the proportion of delayed delivery orders to the total number of orders. The deviation between the actual delivery delay rate and the delivery delay rate predicted by the elastic early warning model in step S04 is calculated. When the deviation exceeds 5%, the model prediction accuracy is deemed to have decreased and needs updating. At this time, weather factor data during the execution period, including rainfall, wind speed, and other meteorological parameters, is obtained from the meteorological data platform, and traffic congestion data, including road congestion index and accident count, is obtained from the traffic management department. These weather factor data and traffic congestion data are combined to form a supplementary feature vector, which is then concatenated with the feature vector composed of the original historical supply disruption data, supplier capacity fluctuation data, and transportation delay data to form a feature vector with expanded dimensions. The parameter update process employs a transfer learning strategy, retraining only the weight matrix of the input layer of the elastic early warning model to adapt to the expanded feature dimensions, while keeping the parameters of the multi-head attention layer and output layer unchanged to retain the learned temporal dependency patterns. The training update minimizes the mean squared error between the predicted results and the actual observations of the expanded feature vectors using the gradient descent algorithm. A learning rate of 0.0001 is set for fine-tuning. Parameter updates are completed when the mean squared error decreases to a convergence threshold below 0.03. This step aims to continuously optimize the early warning model based on actual operational feedback, improving the system's adaptability to environmental changes.

[0053] It should be noted that the key technical ideas of this invention include the following aspects. The first key technical idea is to construct a supply chain knowledge graph using a bidirectional attention disambiguation model. Through bidirectional semantic feature extraction and attention weight calculation, it achieves accurate identification of polysemous words in industrial texts. Compared to traditional rule-based entity recognition methods, this model can dynamically disambiguate based on context, avoiding errors in association caused by semantic confusion, thereby improving the accuracy of supply chain element traceability. The second key technical idea is to introduce the principle of thermodynamic entropy increase into the supply chain scheduling algorithm. It measures the balance of inventory distribution through information entropy and combines it with a simulated annealing process to achieve global optimization search. Compared to traditional heuristic algorithms that are prone to getting trapped in local optima, this method achieves large-scale exploration at high temperatures and fine convergence at low temperatures through a temperature-controlled random acceptance mechanism, significantly reducing computational complexity while ensuring the global optimality of the scheduling scheme. The third key technical approach is to use the Transformer architecture to replace the traditional Long Short-Term Memory (LSTM) network in constructing a resilient early warning model. This model uses a multi-head attention mechanism to capture dependencies at different time scales in parallel. Compared to recurrent neural networks, which can only sequentially process time series data and suffer from long-term dependency decay, this architecture can directly establish connections between any two time steps, improving the ability to identify long-term supply risk patterns. The synergistic effect of these technical approaches lies in the fact that the knowledge graph provides precise entity relationship support for the scheduling algorithm, enabling resource allocation based on complete supply chain lineage tracing. The high-precision prediction of the early warning model provides reliable timing judgments for triggering the scheduling algorithm, while the global optimization capability of the scheduling algorithm ensures the generation of an optimal solution that balances efficiency and balance during early warning response. These three elements form a closed loop from risk identification to knowledge association to intelligent decision-making. Compared to the problem of independent operation and lack of synergy in existing technologies, this invention achieves global intelligent and adaptive optimization of supply chain control.

[0054] It should be noted that this invention also addresses the following technical problem: in multi-source heterogeneous industrial data environments, the differences in data formats of different communication protocols lead to low accuracy in identifying supply chain entity relationships. This invention uses a protocol parsing engine to parse the message structures of OPC UA and MQTT protocol data, extracting semantic information and uniformly converting it into standard format data containing timestamps, device identifiers, parameter types, and values ​​according to predefined mapping rules, providing standardized input for subsequent entity recognition. The bidirectional attention disambiguation model extracts contextual semantic features through a bidirectional long short-term memory network layer. The attention layer calculates the similarity weight between each word and candidate entities in the industrial dictionary, and the output layer combines prior probabilities from the knowledge base for Bayesian fusion, effectively eliminating entity ambiguity in industrial texts, establishing precise correlations between material identifiers, supplier identifiers, production work order identifiers, and transportation order identifiers, and constructing a high-quality supply chain knowledge graph.

[0055] A second aspect of the present invention provides a computer-readable storage medium storing program instructions, which, when executed in a computer, are used to perform the above-described intelligent control method for an industrial supply chain based on an AI vertical domain model.

[0056] A third aspect of the present invention provides an intelligent control system for industrial supply chain based on an AI vertical domain model, comprising the aforementioned computer-readable storage medium. The system is any one of a computer, a server, or a microcontroller. The computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes the program instructions stored in the computer-readable storage medium.

[0057] Specifically, the principle of this invention is as follows: The solution addresses the problem of insufficient accuracy in supply chain disruption prediction. Its principle lies in the fact that the resilient early warning model employs a Transformer architecture's self-attention mechanism to capture the direct dependencies between any positions in the feature vector sequence, overcoming the limitation of traditional recurrent neural networks that can only process time series sequentially. A multi-head attention layer projects the input feature vectors into multiple subspaces and computes attention weights in parallel. Each attention head focuses on the correlation between different time scales and feature dimensions, integrating multi-scale information through a weighted aggregation mechanism, enabling the model to simultaneously capture short-term fluctuations and long-term trends. Positional encoding injects absolute or relative positional information into the time series, ensuring the model understands the temporal relationships of the feature vectors. When the prediction deviation exceeds a threshold, the system quickly adapts to the new data distribution by expanding the feature vector dimension and updating only the input layer weight matrix, avoiding the computational overhead of retraining the entire network. This incremental learning mechanism allows the model to continuously absorb real-time feedback information and gradually correct prediction deviations, aligning with the dynamic characteristics of the supply chain environment.

[0058] The following provides a specific embodiment 1 of the present invention. The specific implementation of steps S01 and S02 in this embodiment 1 is the same as that described above, and will not be repeated in detail here. The specific implementation of other steps is described in detail below.

[0059] In the specific implementation of step S03, the outlier determination formula of the three-standard-deviation principle is expressed as follows:

[0060] ;

[0061] In the formula, The first of the equipment operating parameters Each sample value, the unit of which depends on the parameter type; This is the sampling sequence number, ranging from 1 to the total number of samples, and is dimensionless. The average value of the equipment operating parameters, in units of and . same; The standard deviation of the equipment operating parameters, in units of 1 / 2 and 2 / 3. same; is the standard deviation factor, empirically valued at 3, and dimensionless. The condition is determined when the above inequality holds. These are outliers and will be removed.

[0062] In the specific implementation of step S04, the supply interruption probability value is calculated by the output of the elastic early warning model. This model adopts a converter architecture, and the input layer converts the feature vector... With position encoding The sum is used to form the encoding vector. The formula is expressed as follows:

[0063] ;

[0064] In the formula, It is an encoded vector; The input feature vector contains historical supply disruption data, supplier capacity fluctuation data, and transportation delay data. This is the location encoding vector. After the multi-head attention layer computes the dependencies at different time scales in parallel, the output layer maps the hidden states to supply interruption probability values ​​through a fully connected neural network. ,when It outputs a warning signal in real time. Among them, The preset threshold is set as the probability value corresponding to the 95th percentile based on the statistical distribution of historical supply interruption events, and is dimensionless. This represents the probability of supply disruption within 72 hours, and is dimensionless.

[0065] In a specific implementation of step S05, the formula for calculating the information entropy of the supply chain balance scheduling algorithm constrained by thermodynamic entropy increase is expressed as follows:

[0066] ;

[0067] In the formula, The information entropy of the inventory distribution, expressed in bits; The total number of nodes in the supply chain, dimensionless; The node number is 1 to 1. Dimensionless; For the first Inventory quantity at each node, in units of pieces; The total inventory across all nodes, expressed in units of pieces, is calculated using the following formula: The formula for calculating the state transition probability is as follows:

[0068] ;

[0069] In the formula, Let be the state transition probability, which is dimensionless; The information entropy difference between adjacent states, in bits; The reference information entropy value is equal to the current system information entropy, and the unit is bits. is the normalization coefficient of the Boltzmann constant, with a value of 1, and is dimensionless; This is the current temperature parameter, in Kelvin. The reference temperature value is equal to the initial temperature and is expressed in Kelvin. The temperature decay function is described as follows:

[0070] ;

[0071] In the formula, For the first The temperature parameters after the next iteration are in Kelvin. The number of iterations is dimensionless. This is the initial temperature, typically taken as 1000, and the unit is Kelvin. The temperature decay coefficient is empirically valued at 0.95 and is dimensionless. The algorithm flow includes initializing the temperature parameters. Calculate the initial information entropy based on the initial resource allocation scheme. Enter the iterative loop, randomly generate a new scheme within the neighborhood of the current scheme, and calculate the information entropy of the new scheme. ,like If the new solution is accepted, it will be accepted directly; otherwise, the state transition probability will be used. Decide whether to accept the solution. If accepted, update the current solution and decrease the temperature according to the temperature decay function. When the temperature is below the termination threshold or the number of iterations reaches the upper limit, output the optimal resource allocation solution. The initial information entropy, in bits; The information entropy of the new scheme is expressed in bits.

[0072] In the specific implementation of step S06, the formula for calculating the total carbon emissions across the entire chain is as follows:

[0073] ;

[0074] In the formula, Total carbon emissions across the entire value chain, expressed in kilograms. equivalent; The total number of nodes in the graph neural network is dimensionless. The node number is 1 to 1. Dimensionless; For the first Carbon emissions per node, in kilograms Equivalent. Node carbon emissions are derived from raw material batch data, transportation route data, and production process parameters through graph neural network correlation analysis. Output emission reduction optimization schemes in real time, among which Carbon emission threshold, in kilograms Equivalent, set according to industry standards.

[0075] In a specific implementation of step S07, the formula for calculating the deviation between the actual delivery delay rate and the predicted delivery delay rate is as follows:

[0076] ;

[0077] In the formula, The deviation value is dimensionless. This represents the actual delivery delay rate, which is dimensionless. To predict delivery delay rates, the elastic early warning model calculates the value based on eigenvectors; this value is dimensionless. When... A time-triggered parameter update mechanism feeds weather and traffic congestion data as supplementary feature vectors back to the elastic early warning model. The formula for constructing the extended feature vectors is as follows:

[0078] ;

[0079] In the formula, To expand the feature vector; The original feature vector contains historical supply disruption data, supplier capacity fluctuation data, and transportation delay data; To supplement the feature vector, weather and traffic congestion data are included. Parameter updates employ the gradient descent algorithm, with the loss function being the mean squared error, expressed in the following formula:

[0080] ;

[0081] In the formula, The mean squared error is expressed in units identical to the target value. The sample size is dimensionless. The sample number is a number ranging from 1 to 1. Dimensionless; For the first The actual result for each sample, in units that depend on the prediction target; For the first Prediction results for each sample, in units of Same. When The parameter update is completed when the value decreases to below the preset convergence threshold.

[0082] It should be noted that the variables involved in this invention are explained in detail in Table 1.

[0083] Table 1. Variable Explanation Table

[0084]

[0085] To better understand and implement this invention, the following is a specific application scenario of the invention, Example 2: To verify the effectiveness of the invention, technicians built a test environment to simulate the supply chain control scenario of an electronics manufacturing company. This company has 15 core suppliers, its production line includes 8 injection molding machines and 12 surface mount technology (SMT) machines, its logistics network covers 6 regional warehouse nodes, it processes an average of 420 sales orders per day, and its supply chain involves 45 key materials. Technicians selected operational data from October to December 2024 as the historical dataset, set a week in December 2024 as the test time window, and constructed a supply chain disruption risk scenario for control verification.

[0086] Technicians first deployed multi-protocol data acquisition terminals in the workshops of 15 suppliers. These terminals connected to the programmable logic controllers (PLCs) of the production equipment via the OPC UA protocol, collecting temperature parameters from injection molding machines ranging from 160°C to 210°C, and placement speed parameters from pick-and-place machines ranging from 15,000 to 28,000 pieces per hour. Simultaneously, IoT sensors were accessed via the MQTT protocol to collect raw material inventory data, including 8,500 kg of polycarbonate resin and 125,000 integrated circuit chips. The protocol parsing engine uniformly converted the node identifiers of the OPC UA protocol and the topic-level data of the MQTT protocol into a JSON object representation format, generating standard format data containing timestamps, device identifiers, parameter types, and values. The acquisition cycle was set to 30 seconds. Figure 2 As shown, a total of 86,400 records of equipment operating parameters, 5,280 records of material inventory changes, and 4,320 records of carbon emission monitoring data were collected within 72 hours.

[0087] Technicians constructed a knowledge graph from the collected standard-format data, extracting 18,500 text records containing fields such as material name, supplier name, and work order number from the work order text. The input layer of the bidirectional attention disambiguation model receives 300-dimensional word vectors after word embedding transformation, and the bidirectional long short-term memory network layer extracts contextual semantics through forward and backward scanning. For the polysemous word "polycarbonate," the model calculates its similarity weights with two candidate entities in the industrial dictionary: "chemical raw material polycarbonate" and "optical material polycarbonate," with weights of 0.83 and 0.17 respectively. Combining the prior probability of 0.72 for the "chemical raw material" entity in the knowledge base, and obtaining a posterior probability of 0.91 through Bayesian fusion, the word is disambiguated as a "chemical raw material" entity. Model training uses 4,200 manually labeled samples and 2,800 pseudo-labeled samples generated by remote supervision. The training and validation sets are divided in an 8:2 ratio. After 45 training epochs, the validation set F1 score reaches 0.89, but fails to improve for five consecutive epochs, triggering early stopping. Figure 3 As shown, the completed supply chain knowledge graph contains 45 material nodes, 15 supplier nodes, 8600 work order nodes, 12400 transportation order nodes, and 32800 interconnected edges between nodes.

[0088] Technicians cleaned the standard format data. The historical average temperature parameter of the injection molding machine was calculated to be 185℃ with a standard deviation of 11℃. Using a three-times-standard-deviation principle, 126 outliers (below 152℃ or above 218℃) were removed. The identifiers of 15 suppliers were anonymized using the SHA-256 hash function, mapping supplier names to 64-bit hash values. The cleaned data was written to a log structure merged tree time-series database. The memory buffer table capacity threshold was set to 64MB. When the buffer table reached the threshold, a flush operation was triggered, and the data was batch-dumped to a disk storage table. The disk storage table uses a columnar storage format. The equipment status field was compressed using run-length encoding, compressing 180 consecutive running status records into a running × 180 encoded form, achieving a compression ratio of 3.2. Query operations used a Bloom filter to quickly determine the existence of timestamps. The Bloom filter used five hash functions, keeping the false positive rate below 0.01.

[0089] Technicians constructed a resilient early warning model to predict the risk of supply disruptions in the next 72 hours. They extracted 85 historical supply disruption events from the past 180 days, 2700 daily capacity fluctuation data points from suppliers, and 4500 transportation delay data points from a time-series database. This data was organized into time-series feature vectors, with each time step containing 15 feature dimensions, as shown in Table 2. The resilient early warning model uses a Transformer architecture with eight attention heads. The input layer adds the feature vectors to the positional encoding, and the multi-head attention layer calculates the dependencies at different time scales in parallel. The output layer maps the hidden states to supply disruption probability values ​​through two fully connected neural networks. The preset threshold was determined to be 0.78 based on the 95th percentile of the historical supply disruption event probability distribution. In the simulation scenario, on a Wednesday morning in December, the model detected a supplier's chip capacity fluctuation reaching 35%, and the average transportation delay time increased from the normal 12 hours to 26 hours. The calculated supply disruption probability value for the next 72 hours was 0.82, exceeding the preset threshold of 0.78, and the system output an early warning signal. Figure 4 As shown, the probability of supply disruption increases in the 48 hours before the warning is triggered, gradually climbing from 0.45 to 0.82. Meanwhile, the model predicts a delivery delay rate of 18%, which serves as the benchmark for subsequent effect evaluation.

[0090] Table 2. Feature Vector Dimension Explanation Table

[0091]

[0092] Upon receiving the early warning signal, technicians activated the supply chain balancing scheduling algorithm. They retrieved 328 unfulfilled sales orders from the business system. Real-time inventory data from six regional warehouses showed an uneven inventory distribution: Warehouse 1 had 42% inventory, Warehouse 2 had 8%, and the remaining four warehouses had proportions between 10% and 15%. The algorithm mapped the inventory levels of the six warehouse nodes to the particle distribution of a thermodynamic system, calculating the initial entropy to be 1.85 bits. The algorithm introduced a temperature parameter for simulated annealing optimization, setting the initial temperature to 1000 Kelvin and the temperature decay coefficient to 0.95. During iterative optimization, the high-temperature phase accepted resource allocation schemes that increased the entropy from 1.85 to 2.12 with a probability of 0.68, achieving global search and avoiding local optima. After 320 iterations, the temperature dropped to 50 Kelvin, entering the low-temperature phase, where inferior solutions were accepted with a probability of 0.15 for fine-tuning, ultimately converging to a balanced configuration scheme with an entropy of 2.47 bits. The algorithm outputs a production scheduling plan, including the start-up times for 8 injection molding machines over the next 72 hours. Machine 1 starts at 2 PM on Wednesday, machine 2 starts at 6 PM on Wednesday, and so on. It also outputs a transportation scheduling plan, allocating 3 transport vehicles to transfer 8500 kg of materials from warehouse 1 to warehouse 2, and 6200 kg of materials from warehouse 3 to warehouse 5, balancing inventory distribution across all nodes. Figure 5 As shown, after scheduling optimization, the inventory proportions of the six warehouses were adjusted to 18%, 17%, 16%, 19%, 15%, and 15%, respectively, and the inventory distribution entropy increased from 1.85 to 2.47, with a significant improvement in balance.

[0093] Technicians collected end-to-end carbon emission data based on the production schedule. Raw material batches involved 12 suppliers, transportation routes included 15 road transport routes and 4 rail transport routes, and production process parameters covered 8 categories, such as injection molding temperature and assembly speed. A graph neural network modeled raw material batches as 45 nodes, transportation routes as 19 nodes, and production processes as 8 nodes, establishing 72 edges between nodes to represent material flow relationships. The graph convolutional layer aggregated the carbon emission coefficients of adjacent nodes, and the pooling layer calculated the node contribution weights. The highest carbon emission contribution (0.38) was identified for the road transport route nodes, followed by the injection molding process nodes at 0.27. The output layer summed the total end-to-end carbon emissions, yielding 58,600 kg. The equivalent exceeds the set carbon emission threshold of 50,000 kilograms. Equivalent weight. The system outputs an emission reduction optimization plan, replacing five high-carbon road transport routes with rail transport or new energy vehicle transport. After the replacement, the total carbon emissions across the entire chain are reduced to 48,200 kg. Equivalent, meeting carbon emission constraints. For example... Figure 6As shown, the proportion of carbon emissions in each stage changed before and after optimization. The proportion of the transportation stage decreased from 45% to 32%, while the proportion of the production stage increased from 38% to 49%.

[0094] After 72 hours of executing the production schedule and transportation plan, technicians collected data from the logistics management system showing an actual delivery delay rate of 26%, which deviated by 8% from the predicted 18% in step S04. This exceeded the 5% threshold, indicating the model needed updating. Technicians also obtained rainfall data of 68 mm and wind speed of 12 m / s from the meteorological data platform, and a road congestion index of 7.2 and 3 accidents from the traffic management department. Weather data and traffic congestion data were combined to form a 4-dimensional supplementary feature vector, which was then concatenated with the original 15-dimensional feature vector to form a 19-dimensional extended feature vector. Parameter updates employed a transfer learning strategy, retraining the input layer weight matrix of the elastic early warning model to adapt to the expanded dimensions while maintaining the parameters of the multi-head attention layer and output layer unchanged. The update training used a gradient descent algorithm with a learning rate of 0.0001. After 180 training epochs, the mean squared error decreased from the initial 0.082 to 0.028, falling below the convergence threshold of 0.03, thus completing the parameter update. The updated model re-predicts under similar weather and traffic conditions, adjusting the predicted delivery delay rate to 25%, reducing the deviation from the actual value of 26% to 1%, and significantly improving prediction accuracy.

[0095] The advancements of this invention compared to traditional supply chain management methods are reflected in several aspects. Traditional methods rely on human experience to judge supply risks, resulting in delayed response times and accuracy limited by the knowledge level of schedulers. In contrast, this invention automatically captures risk patterns in historical data through a resilient early warning model and utilizes the multi-head attention mechanism of the Transformer architecture to extract dependencies at different time scales in parallel, achieving forward-looking early warning of supply disruptions within the next 72 hours, thus securing sufficient response time for management decisions. Traditional heuristic scheduling algorithms are prone to getting trapped in local optima, and their computational complexity increases exponentially in large-scale supply chain networks. This invention introduces the principle of thermodynamic entropy increase into scheduling optimization, achieving global exploration at high temperatures and fine convergence at low temperatures through a temperature-controlled simulated annealing process, reducing computational complexity from factorial to logarithmic levels, significantly reducing computational overhead while ensuring global optimality. Traditional supply chain management lacks structured representations of semantic relationships between entities, leading to inefficient material traceability and anomaly localization. This invention constructs a knowledge graph using a bidirectional attention disambiguation model, leveraging contextual semantics and attention weights to achieve accurate entity recognition in industrial texts. It establishes the hierarchical relationships between materials, suppliers, work orders, and orders, providing comprehensive semantic support for scheduling decisions. Traditional carbon emission accounting relies on static statistical methods, failing to dynamically capture the carbon emission propagation effects within the supply chain network. This invention employs graph neural networks to model end-to-end carbon emissions, aggregating carbon emission contributions from adjacent nodes through message passing mechanisms in graph convolutional layers and identifying key high-carbon links through pooling layers, achieving refined dynamic accounting and optimization of the carbon footprint. Traditional control systems lack adaptive learning capabilities, with model performance continuously declining after environmental changes. This invention uses a feedback mechanism to retrain the model using actual operational deviations as supplementary features, employing a transfer learning strategy to update only the input layer parameters while retaining learned temporal patterns, enabling continuous adaptation and optimization of the system to environmental factors such as weather changes and traffic fluctuations.

[0096] 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 changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for intelligent control of industrial supply chains based on AI vertical domain models, characterized in that, Multi-protocol data acquisition terminals are deployed to collect equipment operating parameters, material inventory, and carbon emissions. These data are then converted into standard format data using a protocol parsing engine. A supply chain knowledge graph is constructed to establish entity relationships. The cleaned standard format data is written into a time-series database. Historical supply disruption data, supplier capacity fluctuation data, and transportation delay data are extracted from the time-series database as feature vectors. These vectors are then input into an elastic early warning model to calculate the probability of supply disruption within 72 hours. When the probability of supply disruption exceeds a preset threshold, an early warning signal is output. Upon receiving the early warning signal, a supply chain balance scheduling algorithm based on thermodynamic entropy increase constraints is used to calculate resource allocation schemes and output production and transportation scheduling schemes. Based on the production schedule, a graph neural network is used to calculate the total carbon emissions across the entire supply chain and output emission reduction optimization schemes. Based on the deviation between the actual and predicted delivery delay rates, supplementary feature vectors are fed back to the elastic early warning model for parameter updates. Among them, the supply chain balancing scheduling algorithm constrained by thermodynamic entropy increases maps the inventory of each node in the supply chain to the particle distribution state of the thermodynamic system. The information entropy of the inventory distribution is defined as the negative value of the logarithmic weighted sum of the proportions of each node's inventory to the total inventory, as follows: ； In the formula, The information entropy of the inventory distribution, expressed in bits; The total number of nodes in the supply chain, dimensionless; The node number is 1 to 1. Dimensionless; For the first Inventory quantity at each node, in units of pieces; The total inventory across all nodes, expressed in units of pieces, is calculated using the following formula: .

2. The intelligent control method for industrial supply chains according to claim 1, characterized in that, The protocol parsing engine extracts semantic information from data fields by parsing the message structure of OPC UA protocol data and MQTT protocol data, and converts them into standard format data containing timestamps, device identifiers, parameter types and values ​​according to predefined mapping rules.

3. The intelligent control method for industrial supply chains according to claim 2, characterized in that, Standard format data is stored using JSON object representation.

4. The intelligent control method for industrial supply chains according to claim 3, characterized in that, The steps to construct a supply chain knowledge graph are as follows: First, a two-way attention disambiguation model is used to identify entities and extract relationships from text information in standard format data, and then the association and lineage relationships between material identifiers, supplier identifiers, production work order identifiers, and transportation order identifiers are established.

5. The intelligent control method for industrial supply chains according to claim 4, characterized in that, The input layer of the bidirectional attention disambiguation model receives a sequence of word vectors from industrial text, the bidirectional long short-term memory network layer extracts contextual semantic features, the attention layer calculates the similarity weight between each word and the candidate entity, and the output layer combines the semantic vectors and the prior probabilities of the knowledge base through Bayes' theorem to obtain the disambiguation result.

6. The intelligent control method for industrial supply chains according to claim 5, characterized in that, The training dataset for the bidirectional attention disambiguation model is derived from samples containing entity and relation annotations extracted from historical work order texts. For unannotated samples, a remote supervised learning method is used to automatically generate pseudo-labels. The annotated samples and pseudo-labeled samples are then merged to form the training dataset.

7. The intelligent control method for industrial supply chains according to claim 6, characterized in that, The cleaning process involves removing outliers from equipment operating parameters using the three-standard-deviation principle and desensitizing supplier identifiers.

8. The intelligent control method for industrial supply chains according to claim 7, characterized in that, The time-series database adopts a log structure merge tree architecture, including a memory buffer table and a disk storage table. Newly collected standard format data is first written to the memory buffer table, and when the memory buffer table reaches a preset capacity threshold, it is batch-written to the disk storage table.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores program instructions, which, when executed in a computer, are used to perform the intelligent control method for industrial supply chain based on an AI vertical domain model as described in any one of claims 1-8.

10. An intelligent control system for industrial supply chains based on AI vertical domain models, characterized in that, The system comprises the computer-readable storage medium of claim 9, wherein the system is a computer, the computer-readable storage medium is disposed within the system, and the system is provided with a microprocessor that executes program instructions stored in the computer-readable storage medium.

Images