AI-based full-link data penetration method and platform for manufacturing enterprises

Abstract

The application discloses an AI-based full-link data penetration method and platform for manufacturing enterprises, and belongs to the technical field of industrial data processing, which comprises the following steps: adopting an NLP model capable of online incremental training to convert unstructured requirements into structured documents, generating a unique project identifier through PLM analysis and verification; combining BOM, inventory and supplier portraits to generate a purchase request scheme, and dynamically adjusting production pre-scheduling according to incoming material quality inspection results; collecting working condition data in the production link, optimizing scheduling through an idle window prediction algorithm, and linking equipment maintenance; after completion, aggregating process costs based on a directed acyclic graph, and relying on Internet of Things data to iterate processes after product delivery. The application adopts deep learning and natural language processing technology, combines a rule engine and a multi-objective optimization algorithm, and realizes the penetration and real-time linkage of full-link data such as requirements, procurement, production and operation and maintenance.

Description

Technical Field

[0001] This invention relates to the field of industrial data processing technology, and in particular to an AI-based end-to-end data integration method and platform for manufacturing enterprises. Background Technology

[0002] The daily operations of manufacturing enterprises encompass multiple complex business processes, including customer demand response, material procurement and supply, production planning and scheduling, and equipment operation and maintenance. End-to-end data connectivity technology aims to break down data barriers between different business departments within an enterprise using information technology, unifying and standardizing business documents, execution status, and resource allocation information scattered across different systems. This technology is the fundamental support for realizing the digital transformation and lean management of modern manufacturing enterprises, and its core lies in ensuring the smooth flow and efficient collaboration of business data throughout its entire lifecycle.

[0003] In related technologies, Chinese Patent Application No. CN120278649A discloses a method and system for generating procurement requirements based on multi-source data and AI, including: Step S1, collecting a large amount of procurement data and preprocessing each procurement data; Step S2, storing the preprocessed procurement data and constructing a procurement knowledge graph based on the stored procurement data; Step S3, creating a procurement entity recognition model and a procurement requirement generation model and training them; Step S4, obtaining the input procurement request text, inputting the procurement request text into the procurement entity recognition model to obtain the procurement entity, and matching the associated procurement knowledge from the procurement knowledge graph based on the procurement entity; Step S5, performing compliance verification and supplementation on the associated procurement knowledge; Step S6, inputting the procurement knowledge into the procurement requirement generation model to generate procurement requirement text.

[0004] However, the existing technical solutions mentioned above have the following technical shortcomings: Existing solutions only focus on the single procurement stage, failing to cover the entire process from demand access, compliance verification, production scheduling, and equipment operation and maintenance. This results in severe data silos and a lack of cross-stage connectivity and real-time linkage. Procurement decisions rely on static knowledge graph matching, lacking multi-dimensional supplier evaluation and multi-objective optimization, making it difficult to balance multiple constraints such as quality, delivery time, and cost. Furthermore, the lack of a linkage scheduling mechanism between procurement and downstream production and operation and maintenance makes it impossible to respond to changes in order status and coordinate time conflicts between production and equipment maintenance, easily leading to unplanned downtime or wasted capacity. Simultaneously, the lack of end-to-end data association and storage capabilities prevents the formation of a complete project lifecycle view, failing to meet the needs for business traceability and global decision support. Existing technologies do not form a process self-iterative closed loop for after-sales fault data feedback to PLM, lack a rolling adaptive optimization mechanism for production scheduling with circuit breaker protection, and lack an end-to-end encrypted distributed ledger storage system, failing to simultaneously address the three core requirements of continuous process optimization, self-healing of production anomalies, and reliable traceability of business data. Summary of the Invention

[0005] To address the aforementioned issues, this invention provides an AI-based end-to-end data integration method and platform for manufacturing enterprises. By employing deep learning and natural language processing technologies, combined with a rule engine and multi-objective optimization algorithms, it achieves the integration and real-time linkage of end-to-end data across demand, procurement, production, and operation and maintenance.

[0006] The above objectives can be achieved through the following approach: This AI-based end-to-end data integration approach for manufacturing enterprises includes using an online incrementally trainable NLP model to transform unstructured requirements into structured documents, which are then parsed and verified by PLM to generate unique project identifiers. It combines BOM, inventory, and supplier profiles to generate purchase requisition plans and dynamically adjusts production pre-scheduling based on incoming material quality inspection results. During the production process, operational data is collected, and scheduling is optimized and equipment maintenance is linked through idle window prediction algorithms. After completion, process costs are aggregated based on directed acyclic graphs, and after product delivery, processes are iterated using IoT data.

[0007] Optionally, the step of acquiring customer information and unstructured requirement text, using an NLP model for entity extraction and field alignment to generate a structured requirement form, and conducting online incremental training of the NLP model using manually calibrated data during entity extraction includes: setting a confidence threshold for the NLP model's recognition of core business entities in the unstructured requirement text; obtaining the real-time confidence score of the NLP model when performing entity extraction; if the real-time confidence score is less than the recognition confidence threshold, or if there are unaligned key business fields, triggering an abnormal suspension process, pushing the unstructured requirement text and unaligned fields to a manual verification terminal; receiving the calibration data returned by the manual verification terminal, completing field alignment, generating a structured requirement form, and using the calibration data to immediately complete the online incremental training of the NLP model.

[0008] Optionally, the step of calling the PLM process analysis model to parse the structured requirement form, extract project attribute parameters, and perform compliance verification, and if the verification fails, the form is returned for revision; if the verification passes, a unique project identifier and business detail data are generated, including: parsing the structured requirement form using the PLM process analysis model, extracting customer identifier, product specifications, and expected delivery time, and generating project attribute parameters; performing customer credit rating assessment, product process feasibility analysis, and delivery cycle calculation on the project attribute parameters, and generating compliance verification results; when the compliance verification result is unsuccessful, outputting revision opinions and sending them back to the order entry terminal; when the verification result is successful, matching the approval process template according to the project attribute parameters, and generating a unique project identifier and business detail data.

[0009] Optionally, the step of obtaining the Bill of Materials (BOM), real-time inventory data, and supplier profiles corresponding to the business detail data, calculating and generating an initial purchase requisition plan, initiating purchase requisition approval and price inquiry, and backtracking and recalculating the plan if the approval is rejected or the price inquiry fails includes: parsing the business detail data and extracting material requirement details through the material requirement parsing node; comparing the material requirement details with the real-time inventory data to generate material gap data; obtaining the candidate supplier set corresponding to the material gap data; performing multi-dimensional scoring on the supplier profiles of the candidate supplier set to generate comprehensive supplier score data; when solving the initial purchase requisition plan and the backtracking and recalculating plan, taking the procurement cost, delivery risk, and comprehensive supplier score data as the optimization objectives of a multi-objective optimization function, using the NSGA-III algorithm for iterative optimization, solving for a procurement plan solution set with balanced benefits, and selecting a unique execution plan from the solution set according to preset rules.

[0010] Optionally, the process of generating a purchase order and a process status change event after approval, synchronously updating the business document status, and triggering incoming material quality inspection and production pre-scheduling includes: generating a corresponding purchase order after the purchase approval process is completed, and capturing the status changes of each node in the purchase process to generate the process status change event; if the approval result is rejection, or the inquiry result is determined to meet the preset failure conditions, a purchase requisition anomaly event is triggered, automatically backtracking and retrieving the corresponding stage logs to generate a material replacement prompt or budget correction prompt, and re-entering the corrected business parameters to perform backtracking recalculation. The backtracking and recalculation operation sets the number of retries; when the number of retries is exhausted, it automatically switches to manual intervention. Status node identifiers and business association keys are extracted from the process status change events to generate a status mapping relationship and update the flow status of business documents. When the flow status of the business document changes to "pending receipt," an inspection task is created to trigger incoming material quality inspection, and a production pre-scheduling is generated based on the expected completion time of the inspection task. If the actual arrival time reported by the logistics system is later than the preset receipt deadline, an arrival timeout is determined, an arrival delay anomaly is triggered, and the production pre-scheduling is corrected.

[0011] Optionally, during the production process, equipment parameters, scrap rate, and working hours data are collected in real time. When indicators exceed limits, the idle window prediction algorithm is invoked to update the formal production schedule and simultaneously adjust equipment maintenance tasks. Rolling optimization is performed through cross-system linkage changes, and an iteration upper limit is set. Manual intervention to lock the schedule when limits are exceeded includes: collecting equipment parameters and process completion feedback in real time via an IoT interface; if any indicator is determined to exceed the corresponding preset threshold, a production interruption anomaly is triggered; the list of currently unfinished order tasks and affected equipment is retrieved; the idle window prediction algorithm is invoked to re-predict the equipment idle window; equipment maintenance tasks are dynamically adjusted; and the formal production schedule is updated. Simultaneously, the number of rolling optimization iterations is incremented by one; the impact of the updated formal production schedule on the entire supply chain materials and timing is identified. If it leads to a shift in the material demand window or a secondary material shortage, the corresponding in-transit purchase order is automatically traced back based on the unique identifier of the project, triggering expedited, delayed, or additional order control instructions; the degree of compression of the subsequent production window by the adjusted equipment maintenance task is evaluated. If the latest estimated delivery time exceeds the delivery deadline boundary, or the number of rolling optimization iterations reaches the set iteration limit, the schedule circuit breaker is triggered, the current formal production schedule is locked, automatic rolling optimization is terminated, a delivery warning is generated in reverse, and the process is transferred to manual intervention.

[0012] Optionally, when the indicator exceeds the limit, the idle window prediction algorithm is invoked to update the formal production schedule and simultaneously adjust the equipment maintenance task, including: during the update of the formal production schedule, the original processing time chain and delivery deadline boundary are extracted as scheduling hard constraints, and historical operating status time series data of the target equipment, historical sequence of actual working hours of the process, and environmental disturbance characteristics are collected to construct a multi-dimensional time series feature matrix; the multi-dimensional time series feature matrix is ​​input into a pre-trained long short-term memory network model or a Transformer time series prediction model to predict the dynamic availability probability distribution curve of the target equipment in the future time period; based on the dynamic availability probability distribution curve, with the goals of reducing production interruption risk, improving maintenance fit, and meeting scheduling hard constraints as multi-objective optimization objectives, the basic time period constraints are dynamically adjusted and optimized to determine the final expected idle window of the equipment, and the original equipment maintenance plan is bound to the final expected idle window of the equipment to generate the adjusted equipment maintenance task.

[0013] Optionally, the method further includes: when materials are in the pending receipt state, based on the expected completion time of the inspection task and combined with the preset storage location feature matrix, calling a deep reinforcement learning model to adaptively output the target recommended storage location, and performing a material timing matching integrity check before the formal production schedule is executed, triggering a replenishment warning and suspending the schedule when a locking conflict or material shortage is determined; in finished product shipment management, automatically extracting finished product specifications and shipment time limits, using the reduction of transportation costs and delivery delay risks as a multi-objective function, and using the NSGA-III algorithm to solve the shipment logistics plan.

[0014] Optionally, the step of persistently delivering end-to-end business data and business evidence data to the data lake via a distributed message queue using the project's unique identifier as an index, and synchronizing the data state consistency of multiple heterogeneous systems through a distributed transaction component, includes: extracting semantic features from the data interfaces of heterogeneous systems, converting source system fields into universal standardized semantics based on a semantic automatic alignment algorithm based on similarity calculation; writing key state change events in the end-to-end business into a distributed ledger network after encryption, and generating tamper-proof audit credentials through a consensus mechanism among participating nodes; writing all business documents and task partitions after semantic alignment and evidence verification into the end-to-end data lake using the project's unique identifier as an index, and using a distributed transaction component selected from Flexible Transaction SAGA or TCC to perform final consistency synchronization of cross-system data states, and constructing a directed acyclic graph to generate an end-to-end traceability snapshot.

[0015] Based on the same inventive concept, this invention also provides an AI-based end-to-end data integration platform for manufacturing enterprises. The platform includes: a demand intelligent parsing module, used to acquire customer information and unstructured demand text, employing an NLP model for entity extraction and field alignment to generate a structured demand form; during entity extraction, the NLP model is incrementally trained online using manually calibrated data; and a project compliance management module, used to call a PLM process analysis model to parse the structured demand form, extract project attribute parameters, and perform compliance verification. If the verification fails, the form is returned for revision; if the verification passes, a new requirement form is generated. The system includes a unique project identifier and detailed business data; an optimized purchase requisition generation module, which acquires the Bill of Materials (BOM), real-time inventory data, and supplier profiles corresponding to the detailed business data, calculates and generates an initial purchase requisition plan, initiates purchase requisition approval and price inquiry, and recalculates the plan if approval is rejected or price inquiry fails. Upon approval, a purchase order and process status change event are generated, business document status is updated synchronously, and incoming material quality inspection and production pre-scheduling are triggered; a purchase execution and event publishing module, used to return and isolate defective materials if incoming material quality inspection fails, recalculate pre-scheduling, and initiate alternative material procurement; and a business linkage scheduling module, used for quality inspection... After passing the initial screening, the system transfers the data to the warehouse management system for location allocation, shelving, and kitting verification. Failure to pass the kitting verification triggers a replenishment alert and suspends the schedule. Upon successful verification, the pre-schedule locks resources, converts it to a formal production schedule, and sends it to the manufacturing execution system for workshop production. The equipment maintenance optimization module collects real-time equipment parameters, scrap rates, and working hours during production. When indicators exceed limits, it uses an idle window prediction algorithm to update the formal production schedule and simultaneously adjusts equipment maintenance tasks. It performs rolling optimization through cross-system linkage changes and sets iteration limits; exceeding limits locks the schedule and requires manual intervention. The shipping and financial accounting module is used to handle production completion... After finished products are put into storage, the system automatically plans the target shipment plan and executes the delivery. It summarizes all documents in the entire chain based on the unique identifier of the project and completes the process cost collection and automatic financial accounting through a directed acyclic graph. The after-sales and process iteration module is used to monitor failure risks through the Internet of Things after product delivery. When an anomaly occurs, it generates an after-sales work order and pushes it back to the PLM process analysis model to iterate the process. The full-chain data lake module is used to persistently deliver full-chain business data and business evidence data to the data lake through a distributed message queue using the unique identifier of the project as an index, and completes the data status consistency synchronization of multiple heterogeneous systems through a distributed transaction component.

[0016] Compared with the prior art, the present invention has the following advantages: This invention introduces a natural language processing model to extract entities and align fields in unstructured requirement text, and combines this with a multi-objective optimization algorithm to iteratively optimize procurement solutions, thus achieving an automated workflow from customer demand to material requisition. This approach eliminates the information lag caused by traditional manual data entry and experience-based decision-making, enabling manufacturing enterprises to match procurement resources based on real-time inventory and multi-dimensional supplier profiles, reducing communication costs and compliance risks in the business process.

[0017] This invention utilizes an IoT interface to collect equipment parameters and process completion feedback in real time, and combines this with an idle window prediction algorithm to dynamically adjust equipment maintenance tasks and formal production schedules. This cross-system linkage change mechanism breaks down the information barriers between production planning and workshop execution, enabling the manufacturing system to quickly and adaptively adjust when faced with equipment anomalies or material shortages. This avoids global production stoppages caused by local anomalies and ensures the controllability of order delivery cycles.

[0018] This invention employs a semantic automatic alignment algorithm and a distributed transaction component to handle data interfaces between heterogeneous systems, relying on a directed acyclic graph to complete process cost aggregation and automatic financial accounting. This design ensures the ultimate consistency and immutability of business documents throughout the multi-system flow, solving the long-standing problem of data disconnect between finance and operations faced by manufacturing enterprises, and providing accurate and traceable data support for business decision-making by enterprise management.

[0019] This invention constructs a complete closed loop of IoT after-sales work order reverse push PLM. After delivery, product fault data is automatically fed back to the process end for continuous iteration and optimization of processing parameters. There is no need to manually summarize fault information and manually adjust the process, which continuously reduces the product defect rate and realizes the self-evolution of process throughout the product life cycle.

[0020] This invention designs a dynamically iterative rolling optimization logic for scheduling, which combines equipment idle window timing prediction to adaptively adjust production and maintenance plans; at the same time, it sets a scheduling circuit breaker mechanism, which automatically locks the plan and switches it to manual control when the iteration exceeds the limit or the delivery date exceeds the limit, preventing the scheduling logic from getting stuck in an infinite loop, and taking into account both intelligent scheduling and delivery risk management.

[0021] This invention introduces a distributed ledger network to store end-to-end business change events, generates immutable audit vouchers through a consensus mechanism, and achieves data unification across multiple systems by combining semantic alignment and distributed transactions. It also generates complete project traceability snapshots, meeting the data credibility requirements of enterprises for financial reconciliation, production auditing, and compliance verification.

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

[0023] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a flowchart illustrating the AI-based end-to-end data connectivity method for manufacturing enterprises, as described in an embodiment of the present invention.

[0025] Figure 2 This is a module timing interaction diagram according to an embodiment of the present invention.

[0026] Figure 3 This is a semantic similarity heatmap according to an embodiment of the present invention.

[0027] Figure 4 This is a schematic diagram of the structure of an AI-based end-to-end data connectivity platform for manufacturing enterprises, according to an embodiment of the present invention. Detailed Implementation

[0028] 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 embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0029] Reference Figure 1 One embodiment of the present invention proposes an AI-based end-to-end data integration method for manufacturing enterprises. It adopts deep learning and natural language processing technologies, combined with rule engines and multi-objective optimization algorithms, to achieve the integration and real-time linkage of end-to-end data in demand, procurement, production and operation.

[0030] Module timing interaction diagram as follows Figure 2 As shown, the method in this embodiment specifically includes: S1. Obtain customer information and unstructured requirement text, use an NLP model to extract entities and align fields, and generate a structured requirement form. During the entity extraction process, the NLP model is incrementally trained online by manually calibrating data. Optionally, the step of obtaining customer information and unstructured requirement text, using an NLP model for entity extraction and field alignment to generate a structured requirement form, and in the entity extraction process, performing online incremental training of the NLP model through manual calibration data includes: Set the confidence threshold for the NLP model to identify core business entities in unstructured requirement text; Obtain the real-time confidence score when the NLP model performs entity extraction; If the real-time confidence score is less than the recognition confidence threshold, or if there are misaligned key business fields, an abnormal suspension process is triggered, and the unstructured requirement text and misaligned fields are pushed to the manual verification terminal. The system receives calibration data from the manual verification terminal, completes field alignment, generates a structured requirement sheet, and uses the calibration data to instantly complete online incremental training of the NLP model.

[0031] Specifically, the system receives customer information and unstructured requirement text from external systems. Customer information includes basic data such as customer name and contact information. Unstructured requirement text refers to business requirements described in natural language without a fixed format. An NLP model is used to process the unstructured requirement text. The NLP model refers to a computational architecture that processes human language text using deep neural networks. A Distill-BERT combined with a CRF structure is selected, using 2000 customer requirement texts as training data. Entity annotation is completed using the BIO annotation standard, and training, validation, and test datasets are divided for model training. The model's performance metrics and recognition confidence thresholds are clearly defined. A complete online incremental learning process is configured, with model parameters updated in real time based on manually calibrated data. Lightweight deployment solutions such as model quantization, format conversion, and container encapsulation are also provided, and the details of model training, iterative optimization, and engineering implementation are fully disclosed.

[0032] The processing includes entity extraction and field alignment. Entity extraction identifies lexical fragments representing core business entities from unstructured requirement text. Core business entities refer to key information elements indispensable in the business process. Field alignment maps and fills the extracted core business entities into pre-defined standard data categories. Before performing entity extraction, a confidence threshold for the NLP model's recognition of core business entities in the unstructured requirement text is set. The confidence threshold is the lowest score limit for judging the reliability of the model's extraction results. The confidence threshold is set based on the statistical distribution pattern of previous historical manually annotated records, with a total of 2000 records. The real-time confidence score is obtained when the NLP model performs entity extraction. The real-time confidence score is a quantitative assessment of the model's accuracy in identifying the currently extracted core business entities. The real-time confidence score is obtained by calculating the average predicted probability of each character in the extracted fragment, using the following formula: , in the formula This represents the real-time confidence score. Represents the total number of characters contained in the core business entity; dimensionless. Represents the location predicted by the NLP model. The probability value of a character belonging to the corresponding core business entity.

[0033] The real-time confidence score is compared with the identification confidence threshold. If the real-time confidence score is lower than the identification confidence threshold, or if there are misaligned key business fields during field alignment, an abnormal suspension process is triggered. Misaligned key business fields refer to core business entities that failed to be successfully mapped to the standard data category. An abnormal suspension process pauses the current automated processing flow and transitions to manual intervention. In this state, the unstructured requirement text and misaligned fields are pushed to the manual verification terminal. Misaligned fields have the same meaning as misaligned key business fields. The manual verification terminal is an interactive device for operators to view and correct the model extraction results. The system receives calibration data from the manual verification terminal. This calibration data is manually corrected and represents the mapping relationship between core business entities and standard data categories after manual correction by the operator. Field alignment is completed using the calibration data, and a structured requirement form is generated based on customer information. The structured requirement form is an electronic form that arranges business data according to a fixed format and specifications. Online incremental training of NLP models is performed in real time using calibration data. Online incremental training refers to the process of fine-tuning model parameters using newly acquired calibration data without retraining on all historical data. The parameter update formula is as follows: , in the formula This represents the updated model weights. Represents the current model weights. This represents the learning rate, which is set based on the convergence speed of previous model tuning experiments. This represents the value of the cross-entropy loss function calculated based on the calibration data. This represents the gradient of the loss function with respect to the weights.

[0034] For example, taking customized parts processing as an example, the system receives customer information and unstructured requirement text from an external system. The customer information includes basic data such as customer name and contact information. The unstructured requirement text states that a high-strength steel drive shaft with a length of 50 centimeters needs to be processed. The system uses a natural language processing (NLP) model that can be incrementally trained online to process the unstructured requirement text. The processing includes entity extraction and field alignment. Before performing entity extraction, the system sets the confidence threshold for the NLP model's recognition of core business entities in the unstructured requirement text to 0.85. This confidence threshold is set based on the statistical distribution of previous historical manually annotated records, with a total of 2000 historical manually annotated records. The NLP model performs entity extraction on the unstructured requirement text, extracting the core business entities "high-strength steel" and "drive shaft." The system obtains the real-time confidence score of the NLP model during entity extraction. The drive shaft contains a total of 3 characters, i.e. =3. The probability values ​​predicted by the model for the corresponding character belonging to the corresponding core business entity satisfy the following: Based on the real-time confidence score calculation formula, the total number of characters and the predicted probability data for each character are substituted into the formula. The specific calculation process is as follows: The system will display the confidence score in real time. The model compares the current entity extraction result with the recognition confidence threshold. Since the real-time confidence score of 0.88 is greater than the recognition confidence threshold of 0.85, the model determines that the current entity extraction result is reliable. However, when processing the remaining text length of 50 cm, the model fails to recognize this size parameter, resulting in misaligned key business fields during field alignment. At this point, the system triggers an abnormal suspension, halting the current automated processing flow and switching to manual intervention. The unstructured requirement text and the misaligned fields are pushed to the manual verification terminal. The operator manually corrects and labels 50 cm as a size parameter on the manual verification terminal, generating calibration data containing the mapping relationship between the manually corrected core business entities and standard data categories, and sends it back. The system receives the calibration data from the manual verification terminal, completes field alignment, and fills in high-strength steel, drive shaft, and 50 cm into the pre-defined standard data categories of material, product name, and size, respectively, and generates a structured requirement form based on customer information. Without retraining all historical data, the system uses this calibration data to instantly complete online incremental training of the natural language processing model. The system sets the initial current model weights. Given that the first dimension of the feature vector is a scalar equal to 1.2, set the learning rate for online incremental training. The learning rate is set to 0.01, based on the convergence speed of previous model tuning experiments. The system calculates the cross-entropy loss function value based on calibration data. Then, the gradient of the loss function with respect to the weights can be calculated. It equals 0.5. Based on the online incremental training parameter update formula, the specific current model weights, learning rate, and gradient data are substituted into the formula. The specific calculation process is as follows: The updated model weights are calculated. The value is 1.195. The system uses this instantaneous incremental parameter update to fine-tune the model weights, achieving online incremental training and thus improving the model's recognition and alignment accuracy when processing similar-sized descriptive text.

[0035] S2. Call the PLM process analysis model to parse the structured requirement sheet, extract project attribute parameters and perform compliance verification. If the verification fails, return for revision. If the verification passes, generate a unique project identifier and business detail data. Optionally, the step of calling the PLM process analysis model to parse the structured requirement form, extracting project attribute parameters and performing compliance verification, and returning the form for revision if the verification fails, and generating a unique project identifier and business detail data after successful verification includes: The structured requirement form is analyzed using the PLM process analysis model to extract customer identifiers, product specifications, and expected delivery times, and to generate project attribute parameters. The project attribute parameters are evaluated for customer credit rating, product process feasibility analysis and delivery cycle calculation, and compliance verification results are generated. When the compliance verification result is unsuccessful, a revision suggestion is output and sent back to the order entry terminal; When the verification result is passed, the approval process template is matched according to the project attribute parameters to generate a unique project identifier and business detail data.

[0036] Specifically, the system receives the structured requirement form generated by the preceding process. The system then calls the PLM process analysis model to parse the structured requirement form. The PLM process analysis model is a computational program used to manage product lifecycle data and extract and transform process information according to preset rules. The system parses the structured requirement form using the PLM process analysis model, extracting customer identifiers, product specifications, and expected delivery times. Customer identifiers are coded characters used to uniquely distinguish different customers in the system. Product specifications are a set of technical parameters describing the product's physical characteristics and material requirements. Expected delivery time is the customer's requested delivery deadline. The system combines the customer identifier, product specifications, and expected delivery time to generate project attribute parameters. Project attribute parameters are a comprehensive data set covering the core constraints of the order.

[0037] The project's attribute parameters are assessed through customer credit rating, product process feasibility analysis, and delivery cycle calculation. Customer credit rating assessment involves querying historical transaction records based on customer identifiers and calculating default risk. Product process feasibility analysis compares product specifications with the current workshop equipment's processing capabilities. Delivery cycle calculation calculates production time based on product specifications and assesses whether the expected delivery time is met. Based on the above assessment results, a compliance verification result is generated. The compliance verification result is the final conclusion determining whether the current order meets the execution conditions. The compliance verification result is obtained by calculating a comprehensive compliance score, using the following formula: , in the formula This represents the overall compliance score. Represents customer credit score. This represents the feasibility score of the process. This represents the delivery cycle satisfaction score. This represents the customer's credit weighting coefficient. The weighting coefficient represents the feasibility of the process. This represents the weighting coefficient for the delivery cycle. The weighting coefficient is set based on historical order fulfillment success rate statistics, with a total historical order fulfillment success rate statistic set to 1000. The system sets a compliance scoring threshold. The compliance scoring threshold refers to the minimum score required for an order to proceed to the next stage of the process; this threshold is also set based on historical order fulfillment success rate statistics. A global baseline weight is generated using historical order fulfillment data and the analytic hierarchy process (AHP), with numerical boundaries set to prevent extreme bias. Then, dynamic weights are obtained through real-time offsetting and normalization based on scenario tags such as enterprise type, order attributes, workshop capacity, and customer level. Finally, offline training and online incremental iteration are performed using historical fulfillment data, continuously self-optimizing with the help of AI algorithms to adapt to long-term business changes.

[0038] The overall compliance score is compared to a compliance score threshold. If the overall compliance score is lower than the compliance score threshold, the compliance verification result is deemed unsuccessful. When the compliance verification result is unsuccessful, the system outputs revision suggestions and sends them back to the order entry terminal. The revision suggestions point out the unreasonable aspects of the order parameters and provide text information with modification suggestions. The order entry terminal refers to the interactive interface where business personnel initially submit unstructured requirement text.

[0039] When the overall compliance score is greater than or equal to the compliance score threshold, the verification result is considered passed. When the verification result is passed, the system matches an approval process template based on the project attribute parameters. The approval process template refers to a standardized configuration file with predefined review nodes and workflow paths. Finally, a unique project identifier and detailed business data are generated. The unique project identifier is a globally unique code used to track the project throughout the entire business chain. The detailed business data is a detailed data set containing specific material requirements and process steps.

[0040] For example, continuing with the customized parts processing business scenario, the system obtains a structured requirement form generated from the preceding process, which contains standardized requirement data such as high-strength steel material, drive shaft product name, and 50cm dimension parameters. The system calls the PLM process analysis model to parse this structured requirement form. Through the PLM process analysis model, the system extracts information such as a customer identifier with a specific company number, a product specification of a high-strength steel drive shaft with a length of 50cm, and a desired delivery date of a specified date. This information is then combined to generate project attribute parameters. Subsequently, the system performs customer credit rating assessment, product process feasibility analysis, and delivery cycle calculation based on the project attribute parameters. The system also queries the historical transaction records of this company number to obtain a customer credit score. A feasibility score for the process is obtained by comparing the workshop's processing capabilities. Based on product specifications, production time is calculated and a delivery cycle satisfaction score is obtained. The system sets customer credit weighting coefficients. Process feasibility weighting coefficient Delivery cycle weighting coefficient This weighting coefficient is set based on historical order fulfillment success rate statistics, with the total historical order fulfillment success rate statistics set at 1000. According to the compliance verification result comprehensive compliance score calculation formula defined in the instruction manual, the specific scoring data and weighting coefficients mentioned above are substituted into the formula. The specific calculation process is as follows: The system sets a compliance score threshold of 0.80, which is also based on historical order fulfillment success rate statistics. The system executes a scoring comparison logic; since the overall compliance score of 0.85 is greater than or equal to the compliance score threshold of 0.80, the compliance verification result is deemed passed. When the verification result is passed, the system matches the standard machining approval process template according to the project attribute parameters, ultimately generating a unique project identifier and detailed business data including high-strength steel material requirements and turning process steps, completing the entire process of process analysis and compliance verification.

[0041] S3. Obtain the Bill of Materials (BOM), real-time inventory data, and supplier profile corresponding to the business details data, calculate and generate an initial purchase requisition plan, initiate purchase requisition approval and inquiry, if the approval is rejected or the inquiry fails, backtrack and recalculate the plan, and after approval, generate a purchase order and process status change event, synchronously update the business document status and trigger incoming material quality inspection and production pre-scheduling. Optionally, the step of obtaining the Bill of Materials (BOM), real-time inventory data, and supplier profiles corresponding to the business details data, calculating and generating an initial purchase requisition plan, initiating purchase requisition approval and price inquiry, and recalculating the plan if the approval is rejected or the price inquiry fails includes: The business detail data is parsed, and the material requirement details are extracted through the material requirement parsing node; The material requirements details are compared with the real-time inventory data to generate material shortage data; Obtain the set of candidate suppliers corresponding to the material shortage data; The supplier profiles of the candidate supplier set are scored from multiple dimensions to generate comprehensive supplier score data; When solving the initial purchase requisition scheme and the backtracking recalculation scheme, the purchase cost, delivery risk and the supplier comprehensive score data are used as the optimization objectives of the multi-objective optimization function. The NSGA-III algorithm is used for iterative optimization to obtain a set of purchase schemes with balanced benefits. Then, a unique execution scheme is selected from the set of solutions according to preset rules.

[0042] Specifically, the system receives the business detail data generated in the preceding steps and simultaneously acquires the corresponding Bill of Materials (BOM), real-time inventory data, and supplier profiles. The BOM is a structured list of the components and raw materials required for the product. Real-time inventory data refers to the dynamic quantity record of available materials in the current warehouse. The supplier profile is a multi-dimensional feature model of the supplier built based on historical transaction data. The system parses the business detail data and extracts the material requirement details through the material requirement parsing node. The material requirement parsing node is a calculation unit in the system specifically used to break down material levels and summarize the required quantities. The material requirement details refer to the required quantity and time requirements down to a single material code.

[0043] The material requirements details are compared with real-time inventory data to generate material gap data. Material gap data refers to the quantity to be purchased after subtracting available inventory from the required quantity. The calculation formula is as follows: , in the formula This represents material shortage data. This represents the required quantity in the material requirements list, measured in pieces. This represents the available quantity in the real-time inventory data.

[0044] Obtain the candidate supplier set corresponding to the material shortage data. The candidate supplier set refers to a list of suppliers qualified to supply specific materials. Assess the supplier profiles within the candidate supplier set using multi-dimensional scoring. Multi-dimensional scoring refers to a quantitative evaluation based on multiple aspects such as quality, delivery time, and service. The system generates comprehensive supplier score data. The comprehensive supplier score data refers to the overall evaluation value after summarizing the scores from each dimension. The calculation formula is as follows: , in the formula This represents the overall rating data of the supplier. This represents the score for the quality dimension. This represents the score for delivery time. This represents the service dimension score. Represents quality weight. This represents the weight of the option to be traded. This represents the service weight. The weight coefficient is set based on previous historical supplier evaluation data, with a total of 500 historical supplier evaluation data. The baseline weight is derived from the historical supplier evaluation data, and the weight is dynamically adjusted and the range is verified by combining scenario tags such as enterprise type, material category, procurement urgency, and cooperation relationship. At the same time, procurement fulfillment data is continuously fed back, and the weight is automatically iterated and optimized through offline retraining and online incremental learning to achieve adaptive updates.

[0045] The system calculates and generates an initial purchase requisition plan. The initial purchase requisition plan refers to the material procurement plan calculated for the first time. When solving for the initial purchase requisition plan and subsequent potential backtracking recalculation plans, procurement cost, delivery risk, and supplier comprehensive score data are used as the optimization objectives of a multi-objective optimization function. Backtracking recalculation plans refer to the procurement plans recalculated under abnormal circumstances. Procurement cost refers to the total amount of funds required to acquire materials. Delivery risk refers to the quantified probability that a supplier will fail to deliver on time and in good quality. The multi-objective optimization function refers to a mathematical model that considers multiple conflicting indicators simultaneously. The system uses the NSGA-III algorithm for iterative optimization. The NSGA-III algorithm is a non-dominated sorting multi-objective optimization calculation program based on a reference point mechanism. The system obtains a set of procurement plans with balanced benefits, which refers to the set of alternative plans that achieve balance among multiple optimization objectives. The system selects a unique execution plan from the solution set according to preset rules. Preset rules refer to the pre-defined scheme selection logic, which prioritizes the scheme with the lowest procurement cost. The unique execution plan refers to the procurement plan ultimately selected for actual execution.

[0046] For example, taking the material optimization requisition scenario of customized parts processing business as an example, the system receives the business detail data generated by the preceding process, and extracts the material requirement details of the high-strength steel raw materials required for the drive shaft through the material requirement parsing node, including the customer's required quantity. This equals 100 pieces. The system reads the warehouse management system in real time to obtain the current real-time inventory data and determine the available quantity of this high-strength steel. This equals 40 units. Substituting the above specific quantity data into the formula based on the material shortage data calculation, the calculation process is as follows: The material shortage data was calculated. There are 60 items. The system automatically matches and obtains the corresponding candidate supplier set, including supplier A, supplier B, and supplier C. The system retrieves the historical transaction performance profiles of each supplier, with scores for each dimension ranging from 0 to 1, and the system sets quality weights. Equal to 0.5, option weight Equal to 0.3, service weight The weighting factor is 0.2, set based on historical supplier evaluation data, with a total of 500 historical supplier evaluation data points. Supplier A's quality score Rq is 0.95, delivery time score Rt is 0.90, and service score Rs is 0.85. The specific score data and weighting data for each dimension of Supplier A are substituted into the formula according to the supplier comprehensive score calculation formula. The calculation process is as follows: The comprehensive score data of supplier A was calculated. The value is 0.915. The system uses procurement cost, delivery risk, and supplier comprehensive score data as the optimization objectives of a multi-objective optimization function. It employs the NSGA-III algorithm for multi-generational iterative optimization calculations to generate a set of procurement solutions with balanced benefits. Finally, based on the preset rule of prioritizing the lowest procurement cost, the system selects a unique execution plan from the generated solution set, namely, requesting 60 pieces of high-strength steel from supplier A as the initial procurement plan. During the process of initiating the procurement approval and inquiry, if it is determined that supplier A's quotation exceeds the budget by 12%, meeting the preset failure condition of exceeding the budget by 10%, the system triggers a procurement anomaly event. It automatically backtracks and retrieves the log data of the corresponding stage to generate a budget correction prompt. The corrected business parameters are re-entered into the system to perform a backtracking recalculation. At this time, the system changes the number of retries for the backtracking recalculation operation from 0 to 1, which does not exceed the system's set maximum of 3 retries, thus completing the initial procurement plan solution and anomaly backtracking.

[0047] Optionally, the generation of a purchase order and process status change event after approval, synchronously updating the business document status and triggering incoming material quality inspection and production pre-scheduling includes: After the procurement approval process is completed, a corresponding purchase order is generated, and the status changes of each node in the procurement process are captured to generate the process status change event. If the approval result is rejection, or if the inquiry result is determined to meet the preset failure conditions, a purchase requisition exception event is triggered. The system will automatically backtrack and retrieve the logs of the corresponding process to generate a material replacement prompt or budget correction prompt. The corrected business parameters will be re-entered to perform backtracking recalculation. The backtracking recalculation operation has a set number of retries. When the number of retries is exhausted, the process will automatically switch to manual intervention. Parse and extract the status node identifier and business association key from the process status change event, generate a status mapping relationship, and update the flow status of the business document; When the flow status of the business document changes to the pending receipt status, an inspection task is created to trigger incoming material quality inspection, and a production pre-scheduling is generated based on the expected completion time of the inspection task. If the logistics system reports that the actual arrival time is later than the preset delivery deadline, it determines that the delivery has exceeded the time limit, triggers an arrival delay exception, and corrects the production pre-scheduling.

[0048] Specifically, this involves initiating purchase requisition approval and requesting quotations. Purchase requisition approval refers to the process of authorizing the review of procurement plans. Requesting quotations refers to the action of obtaining material quotations from suppliers.

[0049] If the approval result is rejection, or if the inquiry result is determined to meet preset invalidation conditions, a purchase requisition anomaly event is triggered. An invalidation condition refers to a specific situation that renders the quotation inapplicable; the invalidation condition is set at 10% exceeding the budget. A purchase requisition anomaly event refers to a system record indicating that the procurement process has deviated from the normal path. The system automatically backtracks and retrieves the corresponding stage logs to generate material replacement prompts or budget correction prompts. Corresponding stage logs refer to system files that record detailed information about specific operational steps. Material replacement prompts are system notifications suggesting the use of other similar materials. Budget correction prompts are system notifications suggesting adjustments to the procurement fund amount.

[0050] Re-enter the corrected business parameters to perform a backtracking recalculation. Corrected business parameters refer to the basic business data after manual or system adjustments. The backtracking recalculation operation sets the number of retries. A backtracking recalculation operation refers to re-executing the procurement plan calculation. The number of retries is the maximum number of times the system is allowed to automatically recalculate; this is set to 3. Once the number of retries is exhausted, the process automatically transitions to manual intervention. Manual intervention refers to the step where abnormal situations are handled manually.

[0051] After the procurement approval process is completed, a corresponding purchase order is generated. A purchase order is a formal commercial document confirming the purchase of materials from a supplier. The system captures status changes at each node of the procurement process to generate process status change events. These events record data indicating the transition of the business process from the current stage to the next. The system parses and extracts status node identifiers and business association keys from these events. A status node identifier is a unique code representing a specific stage of the process. A business association key is a common field used to associate different business documents. The system generates a status mapping relationship and updates the flow status of business documents. This mapping relationship refers to the correspondence between status node identifiers and business document statuses. A business document status indicates the current processing progress of a business file within the system. A business document is the business document whose status needs updating. Flow status indicates the current position of the business document in the process. The system synchronously updates the business document status and triggers incoming material quality inspection and production pre-scheduling. Incoming material inspection refers to the process of inspecting the quality of materials delivered by suppliers. Production pre-scheduling refers to the preliminary production task schedule for the workshop. When the status of a business document changes to "pending receipt," the system creates an inspection task to trigger incoming material inspection. The "pending receipt" status indicates that the purchase order has been issued and is awaiting delivery from the supplier. An inspection task refers to the work instructions that guide quality inspectors in performing inspection operations. The system calculates and generates a production pre-scheduling based on the estimated completion time of the inspection task. The estimated completion time refers to the estimated end time of the quality inspection work, calculated using the following formula: , in the formula Represents the start time of the pre-production scheduling; This represents the estimated completion time of the inspection task; This represents the material flow buffer time. The material flow buffer time is set based on the previous workshop logistics measurement data, and the total amount of the previous workshop logistics measurement data is set to 300.

[0052] If the actual arrival time reported by the logistics system is later than the preset receiving deadline, it is considered a timeout. The actual arrival time refers to the real time the goods arrive at the warehouse as recorded by the logistics system. The preset receiving deadline refers to the latest allowed arrival time stipulated by the system. A timeout occurs when the actual arrival time is later than the preset receiving deadline. The system triggers a timeout exception and corrects the production schedule. A timeout exception is a system event that records timeout situations.

[0053] For example, taking the rolling optimization scenario of production scheduling for customized parts processing as an example, the system captures the status change event of the preceding procurement process and updates the flow status of the business document from "pending approval" to "pending receipt". When the flow status of the business document changes to "pending receipt", the system automatically creates an inspection task and triggers the incoming material quality inspection process. The quality inspection department calculates based on the arriving batch and sampling rules to determine the estimated completion time of the inspection task for this batch of high-strength steel. The virtual ephemeris timestamp is 1000. Based on previous workshop logistics measurement data, the workshop logistics system automatically reads the material flow buffer time Tb as 5. This buffer time is dimensionless and based on previous workshop logistics measurement data, with the total amount of previous measurement data set at 300. Substituting the virtual estimated completion time and buffer time data into the production pre-scheduling start time calculation formula, the calculation process is as follows: The start time of the production pre-scheduling corresponding to this project was calculated. The workshop management system generates a formal pre-production schedule based on the ephemeris timestamp 1005. However, during actual execution, the logistics system reports the actual arrival time as the virtual ephemeris timestamp 1020, while the system's preset receiving deadline is the virtual ephemeris timestamp 1010. The system executes time comparison logic, determining that the actual arrival time is later than the preset receiving deadline, thus indicating an arrival delay. The system then automatically triggers an arrival delay exception, calling the scheduling rolling calculation module to recalculate the estimated completion time of the inspection task using the actual arrival time 1020 as the new baseline, thereby dynamically correcting the pre-production schedule and ensuring smooth data flow and dynamic linkage throughout the entire chain.

[0054] S4. If the incoming material fails the quality inspection, the defective material shall be returned, replaced, isolated, and the pre-scheduling shall be recalculated and the procurement of alternative materials shall be initiated. Specifically, the incoming material quality inspection process relies on the inspection tasks generated by the preceding processes. Inspectors conduct full-item testing on the incoming materials according to established quality inspection standards, and simultaneously enter the inspection results into the system. After reading the inspection results, the system makes a judgment. If the material is deemed unqualified, the defective material handling procedure is first executed, and the unqualified materials are uniformly returned or replaced and stored separately to prevent them from entering the production process. Subsequently, the system retrieves the currently generated pre-production schedule and, considering the material shortage situation, recalculates the schedule to complete the pre-scheduling recalculation. Simultaneously, the system searches the Bill of Materials (BOM) and supplier profile data, filters alternative suppliers with the ability to supply similar materials, and initiates the alternative material procurement process.

[0055] S5. After passing quality inspection, the system will be transferred to the warehouse management system to complete the allocation of warehouse locations, shelving and completeness verification. If the completeness verification fails, a replenishment warning will be triggered and the schedule will be suspended. If the verification passes, the pre-schedule will lock the resources and be converted into a formal production schedule and issued to the manufacturing execution system to execute workshop production. Specifically, after the incoming materials pass the quality inspection and the judgment result is qualified, the system automatically transfers the material-related data to the warehouse management system. The warehouse management system combines the estimated completion time of the inspection tasks generated in the previous period and the preset bin feature matrix, calls the deep reinforcement learning model to complete the bin allocation, and completes the material shelving operation according to the allocation result. The bin feature matrix is a standardized data set formed by integrating the carrying capacity, location moving line, and environmental conditions of each storage location in the warehouse. The deep reinforcement learning model is trained based on historical inbound data and can independently match the storage locations that meet the material storage requirements. After the material shelving is completed, the system starts the kit verification. The kit verification is a verification process that compares the material list BOM corresponding to the project, checks the category quantity and arrival time sequence of the currently inbound materials, and determines whether all the materials required for production are complete. If the kit verification does not meet the standard, the system immediately issues a replenishment warning signal, and at the same time suspends the current production pre-scheduling execution status to complete the scheduling suspension operation, waiting for the missing materials to be replenished. If the kit verification passes smoothly, the system locks various production resources such as production equipment, materials, and working hours involved in the production pre-scheduling, converts the production pre-scheduling into a formal production schedule, and then fully distributes the formal production schedule to the manufacturing execution system, and the manufacturing execution system coordinates the workshop production operations.

[0056] S6. During the production process, the equipment parameters, scrap rate, and working hour data are collected in real time. When the indicators exceed the limit, the idle window prediction algorithm is called to update the formal production schedule and synchronously adjust the equipment maintenance tasks. Rolling optimization is performed through cross-system linkage changes and an iteration upper limit is set. When the limit is exceeded, the schedule is locked and manual intervention is required; Optionally, the equipment parameters, scrap rate, and working hour data are collected in real time during the production process. When the indicators exceed the limit, the idle window prediction algorithm is called to update the formal production schedule and synchronously adjust the equipment maintenance tasks. Rolling optimization is performed through cross-system linkage changes and an iteration upper limit is set. When the limit is exceeded, the schedule is locked and manual intervention is required, including: The equipment parameters and process completion feedback are collected in real time through the Internet of Things interface. If it is determined that any indicator exceeds the corresponding preset threshold, a production interruption exception is triggered. The current unfinished order tasks and the list of affected equipment are retrieved, the idle window prediction algorithm is called to re-predict the equipment idle window, the equipment maintenance tasks are dynamically adjusted and the formal production schedule is updated, and the rolling optimization iteration count is incremented by one at the same time; Identify the impact of the updated formal production schedule on the full-link materials and time sequence. If it causes the material demand window to shift or a secondary material gap, the corresponding in-transit purchase order is automatically traced back in reverse according to the project unique identifier, and the purchase urgent, extension, or additional order control instructions are triggered; Assess the extent to which the adjusted equipment maintenance tasks compress the subsequent production window. If the latest estimated delivery time exceeds the delivery deadline boundary, or if the number of rolling optimization iterations reaches the set iteration limit, trigger the scheduling circuit breaker, lock the current formal production schedule, terminate automatic rolling optimization, generate a delivery deadline warning in reverse, and switch to manual intervention.

[0057] Specifically, the system collects equipment parameters and process completion feedback in real time through an IoT interface. The IoT interface refers to the communication channel connecting physical devices and the information system to achieve real-time data transmission. Equipment parameters refer to physical quantity data reflecting the machine's operating status. Process completion feedback refers to data information recording the completion status and time consumption of specific production steps. The system compares the collected data with corresponding preset thresholds. Preset thresholds are the boundary values ​​for determining whether equipment operation or production progress is normal; these thresholds are set based on historical data from 500 sets of normal equipment operation. If any indicator exceeds the corresponding preset threshold, a production interruption anomaly is triggered. A production interruption anomaly is a system alarm status indicating that the production process has stopped for some reason. The determination process is achieved by calculating the indicator deviation rate, using the following formula: , in the formula The deviation rate of the representative indicator. This represents the real-time collected device parameter values. This represents the corresponding preset threshold. When the deviation rate of the indicator is greater than zero, it is determined that the preset threshold has been exceeded.

[0058] After a production interruption is triggered, the system retrieves the list of currently incomplete order tasks and the list of affected equipment. The currently incomplete order tasks refer to the set of production instructions that have not yet been executed. The list of affected equipment refers to the list of machines that are unable to operate normally due to the interruption. The system calls the idle window prediction algorithm to re-predict the equipment idle window. The idle window prediction algorithm is a mathematical model used to calculate the future available time period of equipment. The equipment idle window refers to the time interval during which a machine is not in operation and can be assigned tasks. The calculation formula is as follows: , in the formula This represents the duration of the device's idle window. This represents the total available time. This represents the estimated time required for currently incomplete order tasks. This represents the estimated time required for equipment maintenance tasks.

[0059] The system dynamically adjusts equipment maintenance tasks and updates the formal production schedule based on the re-predicted equipment idle windows. Equipment maintenance tasks refer to regular maintenance work arranged to maintain machine performance. The formal production schedule refers to the final production timetable after resource locking and execution. The system simultaneously increments the rolling optimization iteration count by one. The rolling optimization iteration count refers to the variable that records the number of times the schedule adjustment process is repeated.

[0060] Identify the impact of the updated formal production schedule on end-to-end materials and timing. End-to-end materials and timing refer to the material requirements and timelines throughout the entire business process. If the update causes a shift in the material requirement window or a secondary material shortage, the system automatically traces back to the corresponding in-transit purchase orders based on the project's unique identifier. A shift in the material requirement window refers to the phenomenon where the originally scheduled material usage time is brought forward or delayed. A secondary material shortage refers to the quantity of materials that are shorted again after the initial procurement due to scheduling changes. The project's unique identifier refers to the globally unique code used to track the project throughout the entire business process. In-transit purchase orders refer to purchase contracts that have been issued to suppliers but have not yet been delivered. The system triggers expedited, delayed, or additional order control commands. Expedited, delayed, or additional order control commands refer to system commands that adjust the delivery date or quantity of purchase orders.

[0061] The system assesses the extent to which the adjusted equipment maintenance tasks compress the subsequent production window. The subsequent production window refers to the available production time period after the current point in time. The system determines whether the latest estimated delivery time exceeds the delivery deadline boundary, or whether the number of rolling optimization iterations has reached the set iteration limit. The latest estimated delivery time refers to the product completion date calculated based on the latest schedule; the delivery deadline boundary refers to the latest delivery time allowed by the customer; and the iteration limit refers to the maximum number of times the system is allowed to automatically adjust the schedule, set to 5 based on system computing resources and response time requirements. If any of the above conditions are met, the system triggers a scheduling circuit breaker. A scheduling circuit breaker is a protection mechanism that forcibly stops the system's automatic scheduling calculation. The system locks the current formal production schedule and terminates automatic rolling optimization. Automatic rolling optimization refers to the system's spontaneous scheduling adjustment process. The system generates a delivery date warning and initiates a manual intervention process. A delivery date warning is an alert indicating that an order may not be delivered on time. The manual intervention process refers to the step where complex and abnormal situations are handled manually by operators. This invention features a unique dual protection mechanism of rolling scheduling optimization and circuit breaker. When equipment or materials malfunction, the scheduling is automatically adjusted iteratively. At the same time, it sets a dual circuit breaker condition of iteration limit and delivery deadline, avoiding system lag and order loss caused by infinite loop scheduling, and achieving a balance between production scheduling automation and risk control.

[0062] For example, taking a workshop execution scenario of customized parts processing as an example, the system collects the operating parameters of the processing equipment and the feedback on process completion in real time through an IoT interface. The system obtains the real-time spindle vibration frequency of a lathe during the processing of a drive shaft and determines that it exceeds a preset threshold. The system executes the indicator over-limit comparison logic, comparing the real-time collected equipment parameter values... It equals 120, while the corresponding preset threshold set by the system. It equals 100. According to the indicator deviation rate calculation formula defined in the instruction manual, the specific real-time data and threshold data mentioned above are substituted into the formula. The specific calculation process is as follows: Because the deviation rate of the indicator (0.2%) is greater than zero, the system determines that the indicator exceeds the preset threshold, thus triggering a production interruption anomaly. After triggering the production interruption anomaly, the system retrieves the currently incomplete drive shaft order tasks and the list of affected equipment, and calls the idle window prediction algorithm to re-predict the equipment idle window. The workshop management system reads the total available time of the lathe. Equals 80, estimated time for currently incomplete order tasks. Equals 45, estimated time for equipment maintenance task. The value is 15. Based on the device idle window duration calculation formula defined in the instruction manual, the above-mentioned estimated time data is substituted into the formula. The specific calculation process is as follows: The system dynamically adjusts equipment maintenance tasks and updates the formal production schedule based on the re-predicted equipment idle window, thereby achieving rolling optimization of the production schedule.

[0063] Optionally, when the indicator exceeds the limit, the idle window prediction algorithm is invoked to update the formal production schedule and the equipment maintenance tasks are adjusted synchronously, including: During the process of updating the formal production schedule, the original processing time chain and delivery deadline boundary are extracted as scheduling hard constraints, and the historical operating status time series data of the target equipment, the historical sequence of actual working hours of the process and environmental disturbance characteristics are collected to construct a multi-dimensional time series feature matrix. The multidimensional time series feature matrix is ​​input into a pre-trained long short-term memory network model or a Transformer time series prediction model to predict the dynamic availability probability distribution curve of the target device in future time periods. Based on the dynamic availability probability distribution curve, and with the goals of reducing production interruption risk, improving maintenance fit, and meeting scheduling hard constraints, the basic time period constraints are dynamically adjusted and optimized to determine the final expected idle window of the equipment. The original equipment maintenance plan is then bound to the final expected idle window of the equipment to generate the adjusted equipment maintenance task.

[0064] Specifically, when the system executes the update of the formal production schedule, it first extracts the original processing time chain and delivery deadline boundary of the project through the Manufacturing Execution System (MES) and the IoT interface. The original processing time chain refers to the sequence of time segments in which each process is executed continuously in the initial plan of the order; the delivery deadline boundary refers to the latest delivery time node of the product agreed in the contract. The system uses the original processing time chain and delivery deadline boundary as hard constraints for scheduling, that is, the time boundary restrictions that must be strictly met and cannot be exceeded in the scheduling update algorithm. At the same time, the system collects the historical operating status time sequence data of the target equipment, the historical sequence of actual working hours of the process, and environmental disturbance characteristics through the enterprise database and workshop sensors. The target equipment refers to the specific production machine that has triggered abnormal indicators exceeding the limit or needs maintenance or scheduling adjustment; the historical operating status time sequence data refers to the record of the equipment's start-up, shutdown, fault, standby status, etc., in the past preset period, arranged in chronological order; the historical sequence of actual working hours of the process refers to the actual time spent on each process when processing similar workpieces in the past; the environmental disturbance characteristics refer to environmental parameters that affect the continuity of production, such as the current temperature changes in the workshop, the voltage fluctuations of the power grid, and the backlog of processes above. The system aligns and normalizes the collected historical operational status time-series data, actual work-hour history sequences of processes, and environmental disturbance characteristics to construct a multidimensional time-series feature matrix. The multidimensional time-series feature matrix refers to the numerical matrix formed by synchronously aligning multiple heterogeneous time-series signals on the time axis.

[0065] The system inputs a multidimensional time series feature matrix into a pre-trained Long Short-Term Memory (LSTM) network model or a Transformer time series prediction model. The LTM network model or Transformer time series prediction model refers to a deep learning network architecture capable of capturing long-term dependencies in time series data.

[0066] Long Short-Term Memory (LSTM) networks, based on a gated architecture, effectively address the vanishing gradient problem during training with long-term data. The network consists of an input layer, a normalization layer, multiple LSTM units, fully connected layers, and an output layer. The model input is a multi-dimensional time-series matrix composed of equipment operating status, process time, and environmental parameters. After preprocessing, this matrix is ​​fed into the network, outputting the probability of equipment availability at different future time periods, which is then integrated to form a probability distribution curve. During training, a sliding window is used to create time-series samples, with mean squared error as the loss function. An optimizer iteratively updates parameters, and an early stopping mechanism prevents overfitting. After training, it can predict equipment status in real time, supporting idle window and production scheduling optimization. Transformers, with self-attention as their core, include structures such as position encoding, multi-layer encoders, and fully connected layers. They rely on position encoding to identify time-series order and leverage multi-head attention to capture long-distance data correlations. Their input data and output format are consistent with LSTM, adapting to equipment probability prediction needs. The training process and time-series sample construction rules are similar to LSTM, often employing the AdamW optimizer and regularization strategies to improve performance. This model excels at long-term time series prediction with superior accuracy, but it requires more parameters and computational resources. The two models can be combined for equipment maintenance and scheduling optimization scenarios, depending on the prediction duration.

[0067] The forward propagation calculation of the model predicts the dynamic availability probability distribution curve of the target equipment over future time periods. The dynamic availability probability distribution curve is the line connecting the probability values ​​of the target equipment being able to be put into normal production without failure and downtime at each discrete time point in the future as a function of time.

[0068] Based on the dynamic availability probability distribution curve, and with the multi-objective optimization goals of reducing production interruption risk, improving maintenance fit, and meeting scheduling hard constraints, the system dynamically adjusts and optimizes the basic time period constraints to determine the final expected idle window for equipment. Reducing production interruption risk means prioritizing production tasks during periods with higher dynamic availability probability; improving maintenance fit means aligning equipment maintenance times as closely as possible with critical points of equipment health degradation and probability decay; the basic time period constraint refers to the regular operating time period during which equipment can run; and the expected idle window for equipment refers to the task-free time period determined by the multi-objective algorithm, which is most suitable for inserting equipment maintenance and has the least interference with the production plan. Finally, the system binds the original equipment maintenance plan to the final expected idle window, generating adjusted equipment maintenance tasks and achieving seamless transition between scheduling and maintenance. In calculating the final expected idle window for equipment, the multi-objective optimization needs to assess the probability drop caused by the cumulative decline in equipment health. The evolution of its dynamic availability probability follows the following evolution formula: , in the formula represents the future The dynamic availability probability value of the target device over a given time period. This represents the probability value of the target device being in its initial good condition. This represents the constant of the equipment wear and deterioration rate. This represents the total number of processes completed up to the current moment. Representing the The actual continuous processing time of each process.

[0069] For example, taking the case of a high-precision CNC machine tool in the workshop triggering an excessive scrap rate during the turning of a drive shaft, the above-mentioned operation process of scheduling update and equipment maintenance linkage optimization is reproduced. The system extracts the original processing time chain of the drive shaft project as 8:00 AM to 2:00 PM on a specified day, with the delivery deadline at 8:00 PM on the same day. The system locks these two time periods as scheduling hard constraints in the algorithm. The system automatically retrieves the historical operating status time series data of the CNC machine tool over the past three months, the historical sequence of actual process time consumption, and the current temperature and humidity disturbance characteristics to construct a multi-dimensional time series feature matrix of size 100×8. This matrix is ​​input into a pre-trained long short-term memory network model, and the model outputs the dynamic availability probability distribution curve of the CNC machine tool in the next twelve hours. Among them, the total number of processes completed after the continuous processing of two drive shafts is represented. Equals 2, the actual continuous processing time of each process. and Each is equal to 1.5 hours. The initial probability value of the machine tool being in good working order is known. Equal to 0.98, the equipment wear and deterioration rate constant. It equals 0.04 per hour. Based on the dynamic availability probability evolution formula, substituting the above specific characteristic data into the formula, the specific calculation process is as follows: The dynamic availability probability value for this period is calculated. The probability value decreased to 0.86. Due to the drop in probability value, a multi-objective optimization algorithm was activated, adjusting and optimizing the algorithm with the objectives of reducing production interruption risk and improving maintenance fit. It was found that scheduling maintenance between 2 PM and 3 PM minimized the overall risk and did not exceed the 8 PM delivery deadline. Therefore, the system determined the final expected idle window for equipment to be between 2 PM and 3 PM, and automatically bound the equipment maintenance plan originally scheduled for the next day to this time period, generating an adjusted equipment maintenance task. This successfully completed the adaptive scheduling linkage rolling optimization under abnormal workshop conditions.

[0070] S7. After the finished products are put into the warehouse, the target shipping plan is automatically planned and the shipment is executed. Based on the unique identifier of the project, the entire chain of documents is summarized and the process cost is collected and automatically financial accounting is completed through a directed acyclic graph. Specifically, after the Manufacturing Execution System (MES) completes all workshop production tasks, finished products are transferred to the warehousing area for storage. Once the warehousing system confirms receipt, it sends a status signal to the Shipping and Financial Accounting module. The Shipping and Financial Accounting module reads the finished product specifications and shipping deadlines, using cost reduction and delivery delay risk mitigation as optimization objectives. It then employs the NSGA-III algorithm for iterative optimization, automatically generating a target shipping plan. On-site personnel then complete cargo loading and transportation scheduling according to this plan, and finally execute the finished product delivery. The NSGA-III algorithm is a multi-objective optimization algorithm based on a reference point mechanism, capable of selecting the optimal execution plan under multiple constraints. The system retrieves a pre-defined project unique identifier, a unique code universally applicable throughout the entire business cycle of a single project. Using this project unique identifier, it traverses various business systems, collecting all documents generated across the entire supply chain, including requirements analysis, procurement, execution, production scheduling, and warehousing management, thus centrally summarizing all business documents and corresponding data. The system constructs a directed acyclic graph (DAG) based on the sequential processing steps of the product. The DAG is a topological structure built according to the business flow sequence, where nodes represent each production step, and unidirectional links represent the flow relationships between steps without any loops. Following the flow links in the DAG, the system sequentially calculates the material consumption, labor costs, and equipment maintenance costs generated at each step, gradually completing the process cost aggregation. After completing the process cost aggregation, the system invokes built-in financial accounting rules, uses the aggregated cost data to perform automatic financial accounting, and outputs standardized financial ledgers and accounting results.

[0071] S8. After product delivery, the Internet of Things is used to monitor fault risks. When an anomaly occurs, an after-sales work order is generated and pushed back to the PLM process analysis model to iterate the process. Specifically, after the product is delivered to the customer's site, IoT devices continuously collect and monitor the product's operational status in real time to identify potential fault risks during operation. The Internet of Things (IoT) is a technology system that relies on various sensors and communication networks to remotely collect and transmit terminal data, continuously capturing various status information during product operation. The system continuously compares real-time monitoring data with preset operating standards. Once an operational anomaly is detected, a corresponding after-sales work order is automatically generated. This work order records the service requirements for the product's anomaly and corresponding project information, serving as the basis for after-sales maintenance. The system then pushes the after-sales work order, containing anomaly information and fault characteristics, back to the PLM process analysis model. The PLM model analyzes process parameters and production processes based on the product's entire lifecycle data. After receiving the relevant data, the model conducts process reviews and parameter adjustments based on issues exposed during actual product operation, gradually completing iterative process optimization. It forms a unique two-way closed loop of product after-sales service and process design in the industry. Relying on the reverse flow of terminal IoT fault data to the PLM process analysis model, it continuously corrects product processing parameters and reduces the recurrence of similar defects from the source, which is different from the traditional solution that only produces in one direction and has no process self-optimization.

[0072] S9. Using the unique identifier of the project as an index, the full-link business data and business evidence data are persistently delivered to the data lake via a distributed message queue, and the consistency synchronization of data status of multiple heterogeneous systems is completed through a distributed transaction component.

[0073] Optionally, the step of persistently delivering end-to-end business data and business evidence data to the data lake via a distributed message queue, using the project's unique identifier as an index, and synchronizing the data state consistency of multiple heterogeneous systems through a distributed transaction component includes: Semantic features are extracted from the data interfaces of heterogeneous systems, and the source system fields are converted into universal standardized semantics across the entire chain using a semantic automatic alignment algorithm based on similarity calculation. Key state change events in the entire business chain are encrypted and written into the distributed ledger network, and audit credentials with immutability are generated through the consensus mechanism among participating nodes. Using the project's unique identifier as an index, all business documents and task partitions, after semantic alignment and evidence verification, are written into the full-link data lake. Then, a distributed transaction component selected from the flexible transaction SAGA or TCC is used to perform eventual consistency synchronization of cross-system data status, and a directed acyclic graph is constructed to generate a full-link traceability snapshot.

[0074] Specifically, semantic similarity heatmaps are as follows: Figure 3As shown, a distributed ledger with encrypted evidence storage is adopted as the trusted foundation for end-to-end data. All business change events are encrypted and uploaded to the blockchain, and node consensus generates immutable auditable credentials. Combined with SAGA / TCC distributed transactions, this completely solves the industry pain points of data inconsistency and lack of audit traceability in multi-heterogeneous systems. Semantic feature extraction is performed on the data interfaces of heterogeneous systems. Heterogeneous systems refer to independent software systems built with different underlying architectures or programming languages, and data interfaces refer to the communication boundaries for data exchange between systems. Semantic feature extraction refers to the process of extracting feature vectors representing the business meaning from the names and contents of data fields. The system uses a semantic automatic alignment algorithm based on similarity calculation to convert source system fields into end-to-end universal standardized semantics. The semantic automatic alignment algorithm based on similarity calculation refers to a calculation program that matches fields with the same business meaning by calculating the distance between feature vectors. Source system fields refer to the originally defined data column names in heterogeneous systems, and end-to-end universal standardized semantics refer to the standard data field names uniformly defined throughout the entire business process. The system calculates the semantic similarity between source system fields and end-to-end universal standardized semantics using the following formula: , in the formula Represents semantic similarity. The feature vector representing the source system field. This represents a feature vector with standardized semantics across the entire process. The system sets a similarity matching threshold, which is based on previous historical data interface mapping records, with a total of 1000 historical data interface mapping records. When the semantic similarity is greater than or equal to the similarity matching threshold, alignment is considered successful.

[0075] The system encrypts and writes key state change events from the entire business process to the distributed ledger network. The entire business process refers to business activities spanning from demand to delivery; key state change events refer to records of changes in the state of core nodes in the business process; and encryption refers to the algorithmic operation of converting plaintext data into ciphertext. The distributed ledger network is a decentralized database system maintained and synchronized by multiple independent nodes. It generates immutable audit credentials through a consensus mechanism among participating nodes. Participating nodes are servers in the distributed ledger network responsible for verifying and storing data. The consensus mechanism refers to the algorithmic rules by which participating nodes reach an agreement on data validity. Audit credentials are electronic records used to prove that business operations actually occurred and are immutable.

[0076] The system uses a unique project identifier as its index. A unique project identifier refers to a globally unique code used to track the project throughout the entire business process. The system writes all business documents and task partitions, after semantic alignment and evidence verification, into the end-to-end data lake. Semantic alignment refers to converting source system fields into standardized semantics applicable to the entire process. Evidence verification refers to the process of confirming the authenticity and validity of audit vouchers. All business documents refer to the collection of all electronic forms generated in the business process. Task partitions refer to storage areas where all business documents are divided according to business modules. The end-to-end data lake is a central repository that centrally stores raw data in various formats throughout the entire business process. The system uses either a flexible transaction SAGA or a TCC distributed transaction component to achieve eventual consistency synchronization of cross-system data states. Flexible transaction SAGA refers to a distributed transaction processing mode that splits long transactions into multiple local short transactions and executes them sequentially. The TCC distributed transaction component refers to a computation module that includes three phases—attempt, confirmation, and cancellation—to ensure cross-system data consistency. Cross-system data state refers to the current processing progress of the same business data in different heterogeneous systems. Eventual consistency synchronization refers to the process of ensuring that the data states of all heterogeneous systems reach complete consistency after a certain period of time. Constructing a directed acyclic graph generates a full-link traceability snapshot. A directed acyclic graph refers to a data structure model consisting of nodes and unidirectional edges that does not have closed loops. A full-link traceability snapshot refers to a static view that records the status and flow path of all data at a specific moment in the entire business process.

[0077] For example, taking the project completion and data archiving scenario of customized parts processing as an example, the system obtains a unique project identifier generated in the preceding process, which is used to track the project throughout the entire business chain. The data interfaces of the source system and heterogeneous systems contain fields such as material codes and delivery times. The system extracts semantic features from the data interfaces of these heterogeneous systems. The system uses a semantic automatic alignment algorithm based on similarity calculation to calculate the distance between the feature vectors of the source system fields and the universal standardized semantic fields throughout the entire chain to match fields with the same business meaning. The model then converts the source system fields into corresponding feature vectors. This converts the generalized standardized semantics into corresponding feature vectors. The feature vector corresponding to the known material code is... and The dot product equals 0.88, and their respective moduli satisfy... equal to 1.0 and The value is 0.92. Substituting the above data into the semantic similarity calculation formula defined in the instruction manual, the specific calculation process is as follows: The system sets a similarity matching threshold of 0.90, which is based on previous historical data interface mapping records. The system executes the comparison logic, based on the calculated semantic similarity... If the similarity score is 0.956 or higher than the similarity matching threshold of 0.90, the system determines that the alignment is successful and converts the source system field into a universally standardized semantic for the entire chain. The system encrypts key state change events in the entire chain of business processes and writes them into the distributed ledger network. An immutable audit credential is generated through a consensus mechanism among participating nodes and recorded in the block. Using the project's unique identifier as an index, the system writes all business documents and task partitions, after semantic alignment and evidence verification, into the entire chain data lake. The system uses a distributed transaction component selected from Flexible Transaction SAGA or TCC to achieve eventual consistency synchronization of cross-system data states. The Flexible Transaction SAGA component splits long transactions into multiple local short transactions and executes them sequentially, ensuring that the data states of various heterogeneous systems reach complete consistency after a certain period. Finally, the system constructs a directed acyclic graph to generate a full-chain traceability snapshot, realizing the continuity and traceability of data across the entire chain.

[0078] Optionally, the method further includes: When materials are in the pending receipt state, based on the estimated completion time of the inspection task and combined with the preset storage location feature matrix, the deep reinforcement learning model is called to adaptively output the target recommended storage location. Before the formal production schedule is executed, the material timing matching integrity is checked. If a locking conflict or material shortage is determined, a material replenishment warning is triggered and the schedule is suspended. In finished product shipment management, finished product specifications and shipment deadlines are automatically extracted. The goal is to reduce transportation costs and delivery delay risks as a multi-objective function, and the NSGA-III algorithm is used to solve the shipment logistics plan.

[0079] Specifically, when materials are awaiting receipt, the warehouse management system receives inspection tasks transmitted from previous processes and automatically extracts the estimated completion time of these tasks. Simultaneously, it reads a pre-configured storage location feature matrix. This matrix represents a multi-dimensional state data matrix constructed by numerically translating the shelf height, maximum load capacity, material flow distance to the entrance / exit, and surrounding temperature and humidity parameters of each physical location in the automated warehouse according to spatial coordinate mapping. The estimated completion time of the inspection task and the storage location feature matrix are input features via a network interface and fed into a pre-trained deep reinforcement learning model. This deep reinforcement learning model is a core neural network architecture built on a near-end policy optimization algorithm. It is trained offline using a large amount of historical inbound scheduling logs as environmental samples, learning the optimal mapping between material shelving distance and warehouse space utilization through alternating iterations of the value network and policy network. After receiving the input features, the deep reinforcement learning model performs adaptive inference, and its output layer adaptively outputs the recommended storage locations for shelving. The recommended storage location refers to the spatial coordinate code of the storage location calculated by the model that minimizes the overall logistics handling power consumption and is most conducive to time-series matching. The deep reinforcement learning model is deployed after offline pre-training based on discrete event simulation. The minimum hardware configuration for inference is: 8-core CPU, 16GB memory, and a single NVIDIA T4 GPU. The specific training process is as follows: A digital twin simulator for warehousing and logistics is built, which can use AnyLogic or a self-developed simulation engine based on Python SimPy to completely replicate the warehouse layout, logistics equipment parameters, and warehousing operation rules. Storage location allocation is modeled as a Markov decision process, using the expected completion time of the task, the storage location matrix, and the storage location occupancy status as inputs, and the storage location code as the action. A reward function is constructed by comprehensively considering space utilization, handling power consumption, and time-series matching degree. The PPO algorithm is optimized using a near-end strategy and completed millions of iterations in the simulation environment until the model converges. This method solves the problems of insufficient real-world samples and high on-site trial-and-error costs, improving the accuracy of storage location recommendations. After simulation training, the model outputs recommended storage locations, the system completes material shelving, and initiates the material time-series matching integrity verification process before the formal scheduling is issued.

[0080] Before the formal production schedule is executed, the production scheduling system automatically initiates a material timing completeness verification process. This material timing completeness verification refers to a logical audit procedure that compares the timing overlap between the planned material consumption nodes of each production process and the actual arrival and quality inspection passing nodes of each purchase order to determine whether the materials can meet the continuous production requirements in terms of time. During the verification process, the scheduling system calculates the timing allowance of work-in-process and raw materials before the target process. To accurately quantify this matching tightness and eliminate quantitative discrepancies, the system uses a timing matching balance index for calculation. The timing matching balance is obtained by evaluating the variance of the arrival deviation time of each material and combining it with the rated waiting time for dimensionless processing. The calculation formula is as follows: , In the formula, This represents the balance of timing and matching; the smaller the value, the more synchronized the material arrival time and the higher the completeness of the matching. It represents the total number of key material types involved in the current formal production schedule, which is a dimensionless pure number obtained by parsing the Bill of Materials (BOM). Representing the The estimated completion time for quality inspection of key materials is calculated by retrieving the logistics tracking time of the corresponding purchase order and overlaying it with the historical average of quality inspection time. This represents the arithmetic mean of the estimated quality inspection completion times for all critical materials. It is calculated by summing the estimated quality inspection completion times of all related materials within the current production batch and then dividing by the number of types. get.

[0081] The system executes a logical judgment. If the calculated time-series matching balance $H$ exceeds the system's preset matching tolerance threshold, or if the target storage location is found to be locked by other batches in the underlying database, it is determined that there is a locking conflict or material shortage. The system immediately triggers a material replenishment warning and suspends the schedule through the distributed transaction component, putting the formal production schedule into a pause and waiting state.

[0082] In the finished product shipment management phase, after the finished products are completed and put into storage, the shipment management module automatically extracts the finished product specifications and shipment deadlines through the Enterprise Resource Planning (ERP) system. The finished product specifications refer to the set of physical attributes such as the finished product's geometric dimensions, palletized weight, and special protection categories. The shipment deadline refers to the latest loading and shipping time stipulated in the sales contract. The shipment management module uses the NSGA-III algorithm to solve the shipment logistics plan, with the goal of reducing transportation costs and delivery delay risk as multiple objective functions. The transportation cost refers to the capital expenditure calculated based on the freight route mileage, vehicle tonnage class, and agreed freight rate. The delivery delay risk refers to the probability distribution value of the expected arrival time of the shipped vehicles exceeding the delivery deadline. The NSGA-III algorithm is a non-dominated sorting genetic computation program that maintains population diversity based on a set of uniformly distributed reference points. The algorithm transforms the multi-objective function into an evolutionary process of multi-generational crossover and mutation selection. After iterative optimization through a preset number of generations, it finally solves for a shipment logistics plan with Pareto optimal characteristics, thereby adaptively generating optimal scheduling instructions that include carrier selection, loading schemes, and driving routes.

[0083] For example, taking the material warehousing and finished product shipment scenario of a high-strength steel drive shaft processing project as an example, the system receives a request for 60 pieces of high-strength steel raw materials that have arrived at the factory and are currently in a pending receipt state. The warehouse management system automatically creates an inspection task and extracts the estimated completion time of the inspection task as 3 PM that day. The system simultaneously retrieves the storage location feature matrix of the automated storage and retrieval system, which fully contains the load-bearing capacity and movement distance data of each available storage location in areas A and B. The system inputs the 3 PM time parameter and the storage location feature matrix into a pre-trained deep reinforcement learning model. After policy network inference, the model adaptively outputs the optimal target recommended storage location as "Area A, Row 03, Floor 04, No. 02". Two hours before the formal production schedule is issued to the manufacturing execution system, the production scheduling system automatically triggers a material timing and completeness check. The system analyzes the three key materials required for the production of this batch of drive shafts: high-strength steel, cutting tools, and packaging boxes, i.e., the number of types... The system retrieves logistics and quality inspection logs to obtain the estimated completion time for quality inspection of high-strength steel. The estimated time for tool quality inspection is 1.0 hour. The estimated time for quality inspection of the packaging box is 1.2 hours. The time is 1.4 hours. The system first calculates the arithmetic mean. The time is 1.2 hours. Then, the specific number of categories and time data mentioned above are substituted into the time series matching balance calculation formula. The specific calculation process is as follows: The system's set tolerance threshold is 0.0500. Since 0.0267 is less than the threshold of 0.0500, and the target storage location is not occupied, the system determines that the material sequence integrity check has passed, does not trigger an exception suspension, and the formal production schedule locks the resources and issues them on time. Once the workshop is completed and the finished product drive shaft is stored, the finished product shipping management module is activated. The system automatically extracts standard wooden crates with a length of 50 cm and a weight of 15 kg per crate, with a shipping time limit of 48 hours. The system uses the transportation cost function and the delay risk function as multi-objective optimization objectives and calls the NSGA-III algorithm. After 500 population iterations in the multi-dimensional solution space, the algorithm, after weighing the high cost of expedited logistics and the delay risk of ordinary logistics, solves and outputs the final shipping logistics plan. A specific logistics company is selected to use ordinary vans for shared shipping at 8:00 AM the following morning, achieving the optimal balance between cost and timeliness.

[0084] Based on the same inventive concept, such as Figure 4 As shown, the present invention also provides an AI-based end-to-end data connectivity platform for manufacturing enterprises, the platform comprising: The intelligent demand parsing module is used to obtain customer information and unstructured demand text, and uses an NLP model to extract entities and align fields to generate a structured demand form. During the entity extraction process, the NLP model is incrementally trained online by manually calibrating data. The project compliance management module is used to call the PLM process analysis model to parse the structured requirement form, extract project attribute parameters and perform compliance verification. If the verification fails, it will be returned for revision. If the verification passes, a unique project identifier and business detail data will be generated. The purchase requisition scheme generation module is optimized to obtain the Bill of Materials (BOM), real-time inventory data and supplier profiles corresponding to the business details data, calculate and generate an initial purchase requisition scheme, initiate purchase requisition approval and inquiry, and if the approval is rejected or the inquiry fails, the scheme is recalculated back. After the approval is passed, a purchase order and process status change event are generated, the business document status is updated synchronously, and incoming material quality inspection and production pre-scheduling are triggered. The procurement execution and event release module is used to return or isolate defective materials that fail incoming material quality inspection, recalculate the pre-schedule, and initiate the procurement of alternative materials. The business linkage scheduling module is used to transfer the warehouse management system after the quality inspection is qualified to complete the allocation of warehouse location, shelving and completeness verification. If the completeness verification fails, a replenishment warning is triggered and the schedule is suspended. After the verification is passed, the pre-schedule locks the resources and is converted into a formal production schedule and sent to the manufacturing execution system to execute workshop production. The equipment maintenance optimization module is used to collect equipment parameters, scrap rate and working hours data in real time during the production process. When the indicators exceed the limit, the idle window prediction algorithm is called to update the formal production schedule and adjust the equipment maintenance tasks in sync. Rolling optimization is carried out through cross-system linkage changes and an iteration limit is set. If the limit is exceeded, the schedule is locked and manual intervention is required. The shipping and financial accounting module is used to automatically plan the target shipping plan and execute the shipment after the finished products are put into the warehouse. It summarizes the documents of the entire chain based on the unique identifier of the project and completes the process cost collection and automatic financial accounting through a directed acyclic graph. The after-sales and process iteration module is used to monitor fault risks through the Internet of Things after product delivery. When an anomaly occurs, an after-sales work order is generated and pushed back to the PLM process analysis model to iterate the process. The end-to-end data lake module is used to persistently deliver end-to-end business data and business evidence data to the data lake via a distributed message queue, using the unique identifier of the project as an index, and to complete the consistency synchronization of data status across multiple heterogeneous systems through a distributed transaction component.

[0085] It should be noted that the electrical connections between the various units described above do not necessarily represent direct or indirect connections. Any indirect connection method can be applied to the embodiments of the present invention as long as it achieves the purpose of the present invention. The above descriptions are merely exemplary embodiments of the present invention and should not be construed as limiting the scope of the present invention.

[0086] All equivalent changes and modifications made in accordance with the teachings of this invention are still within the scope of this invention. Those skilled in the art will readily conceive of other embodiments of this invention upon considering the specification and the disclosure of practical truth. This application is intended to cover any variations, uses, or adaptations of this invention that follow the general principles of this invention and include common knowledge or conventional techniques in the art not described herein.

Claims

1. An AI-based end-to-end data integration method for manufacturing enterprises, characterized in that: The method includes: Customer information and unstructured requirement text are obtained, and an NLP model is used for entity extraction and field alignment to generate a structured requirement form. During the entity extraction process, the NLP model is incrementally trained online by manually calibrating the data. The PLM process analysis model is called to parse the structured requirement form, extract project attribute parameters and perform compliance verification. If the verification fails, it is returned for revision. If the verification passes, a unique project identifier and business detail data are generated. Obtain the Bill of Materials (BOM), real-time inventory data, and supplier profiles corresponding to the business details data, calculate and generate an initial purchase requisition plan, initiate purchase requisition approval and inquiry, if the approval is rejected or the inquiry fails, backtrack and recalculate the plan, and after approval, generate a purchase order and process status change event, synchronously update the business document status and trigger incoming material quality inspection and production pre-scheduling. If incoming materials fail quality inspection, the defective materials will be returned, replaced, or isolated, and the pre-scheduling will be recalculated, and the procurement of alternative materials will be initiated. After passing quality inspection, the materials are transferred to the warehouse management system to complete the allocation of storage locations, shelving, and completeness verification. If the completeness verification fails, a replenishment warning is triggered and the schedule is suspended. If the verification passes, the pre-schedule locks the resources, converts them into a formal production schedule, and sends them to the manufacturing execution system to execute workshop production. During the production process, equipment parameters, scrap rate and working hours data are collected in real time. When the indicators exceed the limit, the idle window prediction algorithm is called to update the formal production schedule and adjust the equipment maintenance tasks in sync. Rolling optimization is carried out through cross-system linkage changes and an iteration limit is set. If the limit is exceeded, the schedule is locked and manual intervention is required. After the finished products are put into storage, the system automatically plans the target shipment plan and executes the shipment. Based on the unique identifier of the project, it summarizes the documents of the entire chain and completes the process cost collection and automatic financial accounting through a directed acyclic graph. After product delivery, the Internet of Things is used to monitor fault risks. When an anomaly occurs, an after-sales work order is generated and pushed back to the PLM process analysis model to iterate the process. Using the project's unique identifier as an index, the entire chain of business data and business evidence data is persistently delivered to the data lake via a distributed message queue, and the consistency synchronization of data status across multiple heterogeneous systems is achieved through a distributed transaction component.

2. The AI-based end-to-end data connectivity method for manufacturing enterprises according to claim 1, characterized in that, The process of acquiring customer information and unstructured requirement text, using an NLP model for entity extraction and field alignment to generate a structured requirement form, and conducting online incremental training of the NLP model through manual calibration data during entity extraction includes: Set the confidence threshold for the NLP model to identify core business entities in unstructured requirement text; Obtain the real-time confidence score when the NLP model performs entity extraction; If the real-time confidence score is less than the recognition confidence threshold, or if there are misaligned key business fields, an abnormal suspension process is triggered, and the unstructured requirement text and misaligned fields are pushed to the manual verification terminal. The system receives calibration data from the manual verification terminal, completes field alignment, generates a structured requirement sheet, and uses the calibration data to instantly complete online incremental training of the NLP model.

3. The AI-based end-to-end data connectivity method for manufacturing enterprises according to claim 1, characterized in that, The process involves calling the PLM process analysis model to parse the structured requirement form, extracting project attribute parameters, and performing compliance checks. If the check fails, the form is returned for revision. If the check passes, a unique project identifier and detailed business data are generated, including: The structured requirement form is analyzed using the PLM process analysis model to extract customer identifiers, product specifications, and expected delivery times, and to generate project attribute parameters. The project attribute parameters are evaluated for customer credit rating, product process feasibility analysis and delivery cycle calculation, and compliance verification results are generated. When the compliance verification result is unsuccessful, a revision suggestion is output and sent back to the order entry terminal; When the verification result is passed, the approval process template is matched according to the project attribute parameters to generate a unique project identifier and business detail data.

4. The AI-based end-to-end data connectivity method for manufacturing enterprises according to claim 1, characterized in that, The process of obtaining the Bill of Materials (BOM), real-time inventory data, and supplier profiles corresponding to the business details data, calculating and generating an initial purchase requisition plan, initiating purchase requisition approval and price inquiry, and recalculating the plan if the approval is rejected or the price inquiry fails include: The business detail data is parsed, and the material requirement details are extracted through the material requirement parsing node; The material requirements details are compared with the real-time inventory data to generate material shortage data; Obtain the set of candidate suppliers corresponding to the material shortage data; The supplier profiles of the candidate supplier set are scored from multiple dimensions to generate comprehensive supplier score data; When solving the initial purchase requisition scheme and the backtracking recalculation scheme, the purchase cost, delivery risk and the supplier comprehensive score data are used as the optimization objectives of the multi-objective optimization function. The NSGA-III algorithm is used for iterative optimization to obtain a set of purchase schemes with balanced benefits. Then, a unique execution scheme is selected from the set of solutions according to preset rules.

5. The AI-based end-to-end data connectivity method for manufacturing enterprises according to claim 1, characterized in that, The process of generating a purchase order and a process status change event after approval, synchronously updating the business document status, and triggering incoming material quality inspection and production pre-scheduling include: After the procurement approval process is completed, a corresponding purchase order is generated, and the status changes of each node in the procurement process are captured to generate the process status change event. If the approval result is rejection, or if the inquiry result is determined to meet the preset failure conditions, a purchase requisition exception event is triggered. The system will automatically backtrack and retrieve the logs of the corresponding process to generate a material replacement prompt or budget correction prompt. The corrected business parameters will be re-entered to perform backtracking recalculation. The backtracking recalculation operation has a set number of retries. When the number of retries is exhausted, the process will automatically switch to manual intervention. Parse and extract the status node identifier and business association key from the process status change event, generate a status mapping relationship, and update the flow status of the business document; When the flow status of the business document changes to the pending receipt status, an inspection task is created to trigger incoming material quality inspection, and a production pre-scheduling is generated based on the expected completion time of the inspection task. If the logistics system reports that the actual arrival time is later than the preset delivery deadline, it determines that the delivery has exceeded the time limit, triggers an arrival delay exception, and corrects the production pre-scheduling.

6. The AI-based end-to-end data connectivity method for manufacturing enterprises according to claim 1, characterized in that, During the production process, equipment parameters, scrap rate, and working hours data are collected in real time. When indicators exceed limits, the idle window prediction algorithm is invoked to update the formal production schedule and simultaneously adjust equipment maintenance tasks. Rolling optimization is performed through cross-system linkage changes, and an iteration upper limit is set. Manual intervention to lock the schedule when limits are exceeded includes: The device parameters and process completion feedback are collected in real time through the Internet of Things interface. If any indicator is found to exceed the corresponding preset threshold, a production interruption is triggered. The list of currently unfinished order tasks and affected equipment is retrieved, the idle window prediction algorithm is called to re-predict the idle window of the equipment, the equipment maintenance tasks are dynamically adjusted and the formal production schedule is updated, and the number of rolling optimization iterations is incremented by one. Identify the impact of the updated formal production schedule on the entire supply chain materials and timing. If it causes a shift in the material demand window or a secondary material shortage, automatically trace back the corresponding in-transit purchase order based on the unique identifier of the project and trigger purchase expediting, delay, or additional order control instructions. Assess the extent to which the adjusted equipment maintenance tasks compress the subsequent production window. If the latest estimated delivery time exceeds the delivery deadline boundary, or if the number of rolling optimization iterations reaches the set iteration limit, trigger the scheduling circuit breaker, lock the current formal production schedule, terminate automatic rolling optimization, generate a delivery deadline warning in reverse, and switch to manual intervention.

7. The AI-based end-to-end data connectivity method for manufacturing enterprises according to claim 6, characterized in that, When the indicator exceeds the limit, the idle window prediction algorithm is invoked to update the formal production schedule and the equipment maintenance tasks are adjusted synchronously, including: During the process of updating the formal production schedule, the original processing time chain and delivery deadline boundary are extracted as scheduling hard constraints, and the historical operating status time series data of the target equipment, the historical sequence of actual working hours of the process and environmental disturbance characteristics are collected to construct a multi-dimensional time series feature matrix. The multidimensional time series feature matrix is ​​input into a pre-trained long short-term memory network model or a Transformer time series prediction model to predict the dynamic availability probability distribution curve of the target device in future time periods. Based on the dynamic availability probability distribution curve, and with the goals of reducing production interruption risk, improving maintenance fit, and meeting scheduling hard constraints, the basic time period constraints are dynamically adjusted and optimized to determine the final expected idle window of the equipment. The original equipment maintenance plan is then bound to the final expected idle window of the equipment to generate the adjusted equipment maintenance task.

8. The AI-based end-to-end data connectivity method for manufacturing enterprises according to claim 5, characterized in that, The method further includes: When materials are in the pending receipt state, based on the estimated completion time of the inspection task and combined with the preset storage location feature matrix, the deep reinforcement learning model is called to adaptively output the target recommended storage location. Before the formal production schedule is executed, the material timing matching integrity is checked. If a locking conflict or material shortage is determined, a material replenishment warning is triggered and the schedule is suspended. In finished product shipment management, finished product specifications and shipment deadlines are automatically extracted. The goal is to reduce transportation costs and delivery delay risks as a multi-objective function, and the NSGA-III algorithm is used to solve the shipment logistics plan.

9. The AI-based end-to-end data connectivity method for manufacturing enterprises according to claim 1, characterized in that, The process of persistently delivering end-to-end business data and business evidence data to the data lake via a distributed message queue, using the project's unique identifier as an index, and synchronizing the data status consistency across multiple heterogeneous systems through a distributed transaction component includes: Semantic features are extracted from the data interfaces of heterogeneous systems, and the source system fields are converted into universal standardized semantics across the entire chain using a semantic automatic alignment algorithm based on similarity calculation. Key state change events in the entire business chain are encrypted and written into the distributed ledger network, and audit credentials with immutability are generated through the consensus mechanism among participating nodes. Using the project's unique identifier as an index, all business documents and task partitions, after semantic alignment and evidence verification, are written into the full-link data lake. Then, a distributed transaction component selected from the flexible transaction SAGA or TCC is used to perform eventual consistency synchronization of cross-system data status, and a directed acyclic graph is constructed to generate a full-link traceability snapshot.

10. An AI-based end-to-end data connectivity platform for manufacturing enterprises, applied to the AI-based end-to-end data connectivity method for manufacturing enterprises as described in any one of claims 1-9, characterized in that, The platform includes: The intelligent demand parsing module is used to obtain customer information and unstructured demand text, and uses an NLP model to extract entities and align fields to generate a structured demand form. During the entity extraction process, the NLP model is incrementally trained online by manually calibrating data. The project compliance management module is used to call the PLM process analysis model to parse the structured requirement form, extract project attribute parameters and perform compliance verification. If the verification fails, it will be returned for revision. If the verification passes, a unique project identifier and business detail data will be generated. The purchase requisition scheme generation module is optimized to obtain the Bill of Materials (BOM), real-time inventory data and supplier profiles corresponding to the business details data, calculate and generate an initial purchase requisition scheme, initiate purchase requisition approval and inquiry, and if the approval is rejected or the inquiry fails, the scheme is recalculated back. After the approval is passed, a purchase order and process status change event are generated, the business document status is updated synchronously, and incoming material quality inspection and production pre-scheduling are triggered. The procurement execution and event release module is used to return or isolate defective materials that fail incoming material quality inspection, recalculate the pre-schedule, and initiate the procurement of alternative materials. The business linkage scheduling module is used to transfer the warehouse management system after the quality inspection is qualified to complete the allocation of warehouse location, shelving and completeness verification. If the completeness verification fails, a replenishment warning is triggered and the schedule is suspended. After the verification is passed, the pre-schedule locks the resources and is converted into a formal production schedule and sent to the manufacturing execution system to execute workshop production. The equipment maintenance optimization module is used to collect equipment parameters, scrap rate and working hours data in real time during the production process. When the indicators exceed the limit, the idle window prediction algorithm is called to update the formal production schedule and adjust the equipment maintenance tasks in sync. Rolling optimization is carried out through cross-system linkage changes and an iteration limit is set. If the limit is exceeded, the schedule is locked and manual intervention is required. The shipping and financial accounting module is used to automatically plan the target shipping plan and execute the shipment after the finished products are put into the warehouse. It summarizes the documents of the entire chain based on the unique identifier of the project and completes the process cost collection and automatic financial accounting through a directed acyclic graph. The after-sales and process iteration module is used to monitor fault risks through the Internet of Things after product delivery. When an anomaly occurs, an after-sales work order is generated and pushed back to the PLM process analysis model to iterate the process. The end-to-end data lake module is used to persistently deliver end-to-end business data and business evidence data to the data lake via a distributed message queue, using the unique identifier of the project as an index, and to complete the consistency synchronization of data status across multiple heterogeneous systems through a distributed transaction component.

Images