A manufacturing digital management method and system based on an MES system

Abstract

The application provides a manufacturing industry digitization management method and system based on an MES system, and relates to the technical field of manufacturing industry digitization management. The method comprises the following steps: a global data management and intelligent integration step, which solves the data island problem through multi-protocol adaptation, CDC synchronization and semantic matching; a product family spectrum deep mining step based on a graph neural network, which constructs a product derivation relationship knowledge graph; a real-time data processing and self-adaptive monitoring step based on edge cloud cooperation, which realizes real-time monitoring of equipment state and dynamic threshold optimization; an intelligent supply chain cooperation step based on a large model and multi-objective optimization, which realizes intelligent material source seeking and risk early warning; and a whole-process quality closed-loop control step based on PDCA and digital twinning, which constructs a closed-loop quality management system covering supplier incoming inspection, production process control and finished product inspection. The technical problems of data island, insufficient real-time processing capacity and uneven data quality of the existing MES system are solved, and whole-chain digitization management of the manufacturing industry is realized.

Description

Technical Field

[0001] This invention relates to the field of digital management technology in manufacturing, and in particular to a digital management method and system for manufacturing based on an MES system. Background Technology

[0002] Manufacturing Execution System (MES), as the core information hub connecting the enterprise planning layer (such as ERP) and process control layer (such as PLC, SCADA), aims to optimize and manage the entire production process from order placement to product completion through information transmission. While the deepening of industrialization and intelligent manufacturing has made MES a key infrastructure for achieving production transparency, quality traceability, and efficiency improvement in the manufacturing industry, a series of thorny issues still exist in the actual application and digital management practices of MES systems, severely hindering the effectiveness of the manufacturing industry's digital transformation. This is despite my country having built tens of thousands of smart factories and ranking first globally in the number of "lighthouse factories."

[0003] Existing MES systems face significant data integration challenges during implementation. On one hand, manufacturing enterprises typically run multiple heterogeneous systems such as ERP, PLM, WMS, and SCADA. These systems are often built by different vendors at different times, with varying data formats, communication protocols, and interface standards. MES requires deep integration with these systems, but in actual deployment, the lack of unified compatibility testing and standardized data exchange protocols often leads to problems such as production orders not being automatically synchronized, inconsistent BOM (Bill of Materials) versions, and inconsistent data formats. For example, when implementing MES, a home appliance company failed to adequately test the interface with its existing CRM system, resulting in initial production order data not being automatically synchronized, forcing the project to pause for two weeks for emergency repairs.

[0004] On the other hand, data synchronization and migration between heterogeneous databases has become a technical pain point. Enterprises face the need to migrate from foreign databases to domestic databases, but ensuring data consistency between different database architectures is extremely difficult. Taking vehicle manufacturing as an example, the MES system generates thousands of semi-structured data entries per second, including nested JSON structures. Traditional relational databases struggle to efficiently handle such flexible schema data models. When attempting to synchronize data from document databases like MongoDB to relational databases, technical challenges such as schema mapping disasters, write amplification issues, and lack of time-series out-of-order processing are often encountered, resulting in high data synchronization latency and inconsistent data. Even with CDC (Change Data Capture) tools, data packet loss and out-of-order submissions are still prone to occur under network jitter and peak write scenarios, affecting the real-time monitoring accuracy of the MES system.

[0005] Manufacturing production sites have extremely high requirements for the real-time performance of data acquisition and processing. A modern production line can generate thousands of equipment status logs, sensor point data, and quality inspection information per second, with a daily data volume reaching hundreds of millions to billions of records. However, existing MES systems generally suffer from performance bottlenecks when dealing with this high-frequency, massive-volume time-series data writing pressure.

[0006] Specifically, in equipment status monitoring scenarios, when hundreds of devices simultaneously report data, the main database CPU experiences continuous spikes, and the connection pool is frequently exhausted, causing data write latency to deteriorate from milliseconds to seconds or even minutes. In data query and analysis scenarios, large-scale queries involving sliding window aggregation and cross-time period backtracking often respond slowly, with P99 latency reaching several seconds, making it difficult for production dashboards to display truly real-time waveforms. A certain automotive parts manufacturing company's MES system experienced tag field truncation under peak write pressure in InfluxDB, resulting in inconsistent device identifiers reported by the same device in different shifts, severely impacting the accuracy of quality traceability. This lack of real-time processing capability directly prevents frontline production personnel from obtaining alarm information immediately upon anomalies, often only discovering the problem when defective products have moved to the next process.

[0007] Data is the lifeblood of a MES system, but existing systems have serious shortcomings in data quality management. First, the accuracy of basic data is questionable. Problems such as inaccurate Bills of Materials (BOMs), inconsistent drawing versions, discrepancies between process routes and actual conditions, and inconsistent material coding are common. When MES pushes materials based on an incorrect BOM or assigns work processes based on outdated process routes, the chaos on the production floor is predictable. In the early stages of implementing MES, an electronics manufacturing company suffered a serious loss due to incorrect material coding, resulting in the misplacement of chips worth 800,000 yuan.

[0008] Secondly, real-time data acquisition faces challenges related to equipment compatibility. Production equipment from different eras and manufacturers within the workshop uses varying communication protocols, and some older equipment even lacks data output interfaces. Companies are forced to invest additional costs in installing sensors and acquisition terminals, but even then, the quality of the collected data remains difficult to guarantee, frequently resulting in data loss, outliers, and misaligned timestamps. One precision machinery manufacturer invested an additional 1.5 million yuan to install sensors to collect data from its aging equipment, barely achieving networked monitoring of all equipment in the workshop.

[0009] Furthermore, the data standardization is low. Workers in different shifts may describe and record the same production anomaly in completely different ways. This "dialectal" data cannot be effectively integrated and utilized by the MES system, leading to distorted statistical analysis and ineffective decision support.

[0010] The end users of MES systems are frontline operators and managers, but existing systems generally have flaws in user experience design, leading to the awkward situation where "the system is implemented, but efficiency actually decreases." Many companies, when promoting digitalization, simply scan and input paper work standards into the system, presenting them as PDFs or static documents on workstation tablets. To "demonstrate digital management," workers need to click "I have viewed the screen" after each step, increasing the number of steps and confirmations, thus actually increasing their workload.

[0011] A deeper problem lies in the fact that truly useful information often resides in the minds of experienced employees, while standard work documents fail to extract the key points. When the system merely displays "documents nobody wants to see," it naturally struggles to gain the approval of frontline employees. Young workers operate the system while referring to their handwritten notes, while veteran employees rely solely on experience, rendering the system a dispensable decoration. This "content failure" directly leads to inaccurate data sources for the MES system—workers passively report work and falsify data entries, causing a severe disconnect between the system's data and actual on-site conditions.

[0012] Furthermore, the lack of training and promotion mechanisms exacerbates this problem. Many companies only conduct a single, rote training session before the system goes live, lacking differentiated guidance for different positions and ongoing on-site support. Employees have no one to turn to when they encounter operational difficulties, and eventually abandon the system.

[0013] For large manufacturing groups, collaborative management across multiple production bases and legal entities is a necessity, but existing MES systems lack sufficient support in this area. Different factories may deploy MES systems from different vendors, with varying data formats, business processes, and management standards, making it difficult to obtain a unified view of production operations at the group level. Even when using a system from the same vendor, cross-factory material allocation, capacity sharing, and order collaboration require complex custom development, and often suffer from data delays and unstable interfaces, affecting business continuity.

[0014] The problems are even more pronounced in globalized operations. MES systems have varying levels of support for multiple languages, currencies, time zones, and tax systems. Some localized systems face compliance risks when deployed overseas, while international systems often lack sufficient support for the OEM model and multi-level supplier collaboration common in Asia.

[0015] MES (Manufacturing Execution System) implementation is a complex systems engineering project involving technology, management, processes, and personnel, resulting in a high failure rate. From the initial planning stage, many enterprises have vague goals and unclear needs, only vaguely proposing macro-level objectives such as "improving efficiency" and "achieving digitalization," without specifying the concrete problems to be solved or the key performance indicators (KPIs). Requirements analysis remains superficial, and project teams fail to engage deeply with operators and team leaders on the shop floor, leading to a severe disconnect between system design and actual business processes.

[0016] In terms of system selection, companies often fall into two extremes: one is "price-only," choosing low-priced products with poor industry compatibility; the other is "brand-only," blindly pursuing international brands, resulting in overly complex and rigid systems with high subsequent customization and maintenance costs. From a project management perspective, the lack of a strong project control mechanism, difficulties in cross-departmental coordination, and a lack of communication mechanisms lead to repeated project delays.

[0017] More importantly, companies don't pay enough attention to the "human" factor. Employees' fear of the unknown system, resistance to being monitored, and perception of increased workload, if not effectively addressed, can escalate into passive resistance, data falsification, and other "soft resistance" behaviors. The implementation of an MES system is not only a technological change, but also a management and cultural change, a fundamental aspect that current methods often overlook.

[0018] Compared to large enterprises, small and medium-sized manufacturing enterprises (SMEs) face more severe challenges in MES application and digital transformation. They generally suffer from a predicament of "unwillingness to transform, fear of transformation, inability to transform, and lack of personnel to transform." Some enterprises' understanding of digital transformation is limited to purchasing equipment and implementing systems, lacking systematic and long-term planning. High cost pressures, difficulties in technology adaptation, and a shortage of professional talent have become key bottlenecks restricting the transformation of SMEs. Even when MES projects are launched, the lack of professional IT teams and multi-skilled personnel often leads to system selection errors, loss of control during implementation, and difficulties in subsequent maintenance.

[0019] In summary, existing digital management methods and systems for manufacturing based on MES systems face numerous unresolved issues regarding data integration, real-time processing, data quality, human-computer interaction, system collaboration, project implementation, and adaptation to SMEs. These problems are intertwined and mutually reinforcing, severely hindering the full realization of the value of MES systems. Therefore, providing a novel digital management method and system for manufacturing that effectively addresses these issues has become a pressing technical challenge for those skilled in the art.

[0020] The information disclosed in this background section is intended only to enhance the understanding of the overall background of the invention and should not be construed as an admission or in any way implying that the information constitutes prior art known to those skilled in the art. Summary of the Invention

[0021] The purpose of this invention is to provide a new digital management method and system for manufacturing that can solve several problems existing in current digital management methods and systems for manufacturing.

[0022] To achieve the above objectives, the present invention provides the following solution: A digital management method for manufacturing based on an MES system includes the following steps: Comprehensive Data Governance and Intelligent Integration: A data hub encompassing multiple heterogeneous data sources is constructed, including at least production equipment, Enterprise Resource Planning (ERP), Product Lifecycle Management (PLM), Warehouse Management System (WMS), and Supply Chain Management (SCM). A multi-protocol compatible interface cluster is deployed to collect heterogeneous data from production equipment of different brands and eras in real time. For older equipment lacking digital interfaces, multi-functional data acquisition terminals integrating vibration, temperature, and current sensors are installed. Analog signals from the equipment are acquired through analog-to-digital conversion, and the collected data is converted to a standard format using an OPC UA protocol conversion module. A heterogeneous database synchronization mechanism based on Change Data Capture (CDC) is established, capturing and writing data to the target database in real time when changes occur in the source database. A Merkle Tree-based data consistency verification mechanism is also established, ensuring eventual consistency of cross-system data synchronization through periodic reconciliation and automatic compensation for discrepancies. For the issue of inconsistent material coding between different systems, a semantic similarity matching engine based on a BERT pre-trained model is constructed to automatically calculate the semantic similarity of material descriptions in different coding systems, generate candidate mapping relationships, and implement them after manual confirmation, forming a unified master data view. Product family tree deep mining based on graph neural networks: Product design change history, BOM (Bill of Materials) discrepancies, and process route change records are extracted from the Master Data Management (MDM) and Product Lifecycle Management (PLM) systems. A product family tree knowledge graph is constructed using the graph database Neo4J. This knowledge graph consists of product nodes and edges representing derivative relationships. New generation products developed based on entirely new design concepts serve as parent nodes; second generation products formed by partial design changes to the first generation products serve as child nodes; and third generation products formed by further design changes to the second generation product nodes serve as grandchild nodes. A graph neural network (GNN) is used to embed representation learning into the product family tree knowledge graph, generating feature vectors for each product node. Potential reusable design modules and common issues across generations are automatically identified through node similarity calculations. Each product node is associated with its entire lifecycle quality traceability data, process documents, change history, and customer complaint records, constructing a holistic product profile. Real-time data processing and adaptive monitoring through edge-cloud collaboration: Edge computing nodes are deployed on the production site to preprocess collected equipment status data, process parameters, and quality inspection data. This includes outlier removal based on the isolated forest algorithm, multi-source timestamp calibration based on the NTP protocol, and data compression based on differential coding. Sliding window aggregation analysis is performed using an Apache Flink-based streaming computing engine to calculate in real-time the overall equipment efficiency (OEE), the process capability index (Cpk) of key process parameters, and the defect rate. The formula for calculating the process capability index (Cpk) is as follows:

[0023] Where USL is the upper specification limit, LSL is the lower specification limit, μ is the process mean, and σ is the process mean; the process mean and process standard are calculated based on real-time sampled data within an adaptive sliding window, and the window size adjustment formula is:

[0024] Where W is the adjusted sliding window size, W base The base window size, α is the adjustment factor, and σ current Let σ be the current process standard deviation. target The target process standard deviation is set; when the Cpk value is detected to be lower than the preset threshold or the quality defect rate exceeds the control limit, an alarm is immediately triggered and sent to the relevant workstation terminals, management personnel's mobile terminals and workshop dashboards via multimodal push; Intelligent supply chain collaboration based on large-scale models and multi-objective optimization: Upon receiving a production order, the Bill of Materials (BOM) is automatically parsed. A large-scale pre-trained language model based on the Transformer architecture is integrated to standardize the semantics of material descriptions, enabling automatic material classification, key parameter extraction, and synonym recognition. The system then uses supplier product catalog data from a digital supplier management platform for intelligent matching and sourcing. Combining supplier historical quality performance KPIs, on-time delivery rates, price factors, and ESG (Environmental, Social, and Governance) ratings, a non-dominated sorting genetic algorithm (NSGA-III) with an elitist strategy is employed for multi-objective optimization. The optimization objectives are to minimize procurement cost (C), minimize quality risk (R), and maximize on-time delivery rate (D). This generates a Pareto optimal solution set as a preferred supplier list. The quality risk (R) is dynamically assessed based on supplier historical quality performance data using a Bayesian network model. The risk assessment formula is as follows:

[0025] in, P ( F i Let be the probability of the i-th type of quality failure event. I iThe impact of the i-th type of quality failure event is determined; production plans, material demand forecasts and inventory status are shared in real time through the supply chain collaboration platform, and purchase order generation and payment processes are automatically triggered based on smart contract technology to drive suppliers to carry out JIT just-in-time delivery; End-to-end quality closed-loop control based on PDCA and digital twins: based on the international quality management standard ISO The 9001 methodology and the PDCA cycle are used to construct a full-chain digital quality management system covering supplier incoming material inspection, production process quality control, and finished product outgoing inspection, while simultaneously building a digital twin model of the production process. In the planning phase (P), historical quality data is used to predict key quality control points and set dynamic control targets via a Long Short-Term Memory (LSTM) network. In the execution phase (D), a machine vision online inspection system integrating high-resolution industrial cameras and multispectral imaging equipment acquires product images in real time, and uses an attention-based visual Transformer model for defect identification, supporting real-time detection and classification of various defect types such as scratches, dirt, damage, burrs, color differences, and dimensional deviations. In the inspection phase (C), quality data is automatically summarized and a dynamic quality control chart based on the Exponentially Weighted Moving Average (EWMA) is generated to monitor process stability in real time. In the improvement phase (A), when a systematic quality deviation is detected, the potential risks of similar and derivative products are traced through the product family tree knowledge graph and graph neural network model. Root cause analysis and improvement measure simulation verification are performed using the digital twin model, automatically triggering design changes or process adjustments, and recording the improvement measures and their effects to form a closed-loop improvement knowledge base.

[0026] Optionally, the comprehensive data governance and intelligent integration further includes: establishing a data quality management system, setting data quality dimensions including completeness, accuracy, consistency, timeliness, and uniqueness, scanning data in the data lake through scheduled tasks, marking, alerting, or automatically repairing data that does not conform to the rules based on preset rules; for outliers in the collected data, in addition to using the isolated forest algorithm, a density-based clustering algorithm DBSCAN is also used for secondary verification, and the identified outliers are traced back to their source, distinguishing three different causes: sensor failure, communication interruption, and production process abnormality, and triggering equipment maintenance work orders, network diagnostic processes, or quality warnings respectively; establishing a data lineage tracing mechanism, generating a data lineage map by parsing the data flow path and transformation logic, supporting end-to-end traceability from the final report to the original data source, ensuring data interpretability and problem location efficiency.

[0027] Optionally, the in-depth mining of product genealogy based on graph neural networks further includes: when product quality issues or customer complaints occur, the system automatically extracts the feature vector of the problematic product, performs similarity retrieval in the product genealogy knowledge graph through a graph neural network model, identifies all derivative products with the same parent, child, or grandparent relationship as the problematic product, and predicts its potential risk probability based on historical quality data, pushing differentiated early warning information; the early warning information includes risk level, possible failure modes, suggested stricter sampling inspection schemes, and solutions to similar historical problems; at the same time, the system identifies the chain reaction of design changes through correlation analysis. When the design of a parent product changes, the system automatically assesses the impact on all child and grandchild products, pushes design collaboration notifications, and avoids quality problems caused by design changes not being synchronized to derivative products.

[0028] Optionally, the edge-cloud collaborative real-time data processing and adaptive monitoring further includes: a dynamic optimization mechanism for monitoring thresholds based on reinforcement learning, which learns the correlation between historical alarm data and actual faults through a deep Q-network (DQN) model, automatically optimizing the alarm thresholds of each monitoring indicator, and minimizing the false alarm rate while ensuring no missed alarms; when a continuous downward trend in Cpk or periodic fluctuations in the quality defect rate are detected, potential root causes are identified through a time-series causal inference algorithm, including factors such as raw material batch fluctuations, changes in environmental temperature and humidity, wear of equipment tools, and changes in operators, and pushed to process engineers for confirmation; the multimodal push methods include push notifications from industrial tablets at workstations, vibration alerts from wearable devices, sound and light alarms in the workshop, and mobile APP notifications, automatically selecting the optimal push channel according to the alarm level and personnel role.

[0029] Optionally, the intelligent supply chain collaboration based on large models and multi-objective optimization further includes: integrating external data sources, including macroeconomic data, raw material price indices, logistics freight rate indices, meteorological data, and geopolitical risk indices; and using the time series forecasting model Prophet and the machine learning classification model XGBoost to dynamically assess and warn of supplier bankruptcy risk, raw material shortage risk, logistics disruption risk, and delivery delay risk. The risk warning level is calculated comprehensively based on the probability of occurrence and the degree of impact, using the following formula:

[0030] Among them, AlertLevel is the risk warning level, with a value of an integer from 1 to 5, β is the scaling factor, P is the probability of risk occurrence, and S is the degree of risk impact. When the risk level exceeds the preset threshold, the emergency plan is automatically triggered, including recommending alternative suppliers, adjusting safety stock levels, initiating alternative material certification processes, or changing production plans. At the same time, a reliable traceability platform for the supply chain is built based on blockchain technology to record information on the entire process of materials from raw material procurement, production and processing, quality testing to logistics and distribution, ensuring the immutability and traceability of supply chain data.

[0031] A manufacturing digital management system based on MES includes: The data hub and governance module is used to perform full-domain data governance and intelligent integration. It includes a multi-protocol adaptation interface cluster, an Apache Kafka-based distributed message queue unit, a unified data lake storage unit, a heterogeneous database synchronization unit, and a data quality governance unit. The multi-protocol adaptation interface cluster supports Modbus, OPC UA, Profinet, EtherCAT, CANopen, and MQTT industrial protocols and JDBC / ODBC database interfaces, enabling interconnection with different brands of equipment and heterogeneous information systems. The heterogeneous database synchronization unit adopts a CDC-based synchronization mechanism and integrates a Merkle Tree-based data consistency verification engine. The data quality governance unit has a built-in data quality management rule base, supporting the configuration and execution of rules for integrity, accuracy, consistency, timeliness, and uniqueness. It also includes a BERT-based semantic similarity matching engine to solve material coding inconsistencies. The unified data lake storage unit integrates a relational database, a time-series database InfluxDB, a graph database Neo4j, and a document database MongoDB, used to store structured business data, equipment time-series data, product genealogy knowledge graphs, and unstructured documents, respectively. The Product Genealogy Intelligent Engine module interfaces with the Master Data Management System (MDM) and the Product Lifecycle Management System (PLM) to extract product design change history and BOM differences, construct and visualize the product genealogy knowledge graph, including a graph database storage unit, a graph neural network model training unit, and a similarity retrieval unit. The graph neural network model training unit uses a graph convolutional network (GCN) or a graph attention network (GAT) to learn embedded representations of product nodes and their derived relationships. The similarity retrieval unit supports quickly retrieving related derived products based on feature vector similarity from any product node and displays their quality traceability data, process documents, and change history. The edge-cloud collaborative real-time monitoring module is deployed on the edge computing node cluster and cloud-based centralized management platform at the production site. It includes a data preprocessing unit, a streaming computing unit, an adaptive monitoring threshold optimization unit, and a multimodal alarm push unit. The data preprocessing unit integrates the Isolation Forest algorithm and the DBSCAN algorithm for outlier detection and removal. The streaming computing unit, based on the Apache Flink engine, supports sliding window aggregation analysis and Complex Event Processing (CEP), and calculates the overall equipment efficiency (OEE), process capability index (Cpk), and quality defect rate in real time. The adaptive monitoring threshold optimization unit dynamically optimizes alarm thresholds based on the Deep Q-Network (DQN) model. The multimodal alarm push unit supports integrated push notifications from workstation industrial tablets, wearable devices, workshop audible and visual alarms, and mobile apps. The intelligent supply chain collaboration module interfaces with Supplier Management System (SRM), Enterprise Resource Planning (ERP), and Supply Chain Finance (SPL) systems. It includes an intelligent material sourcing unit, a multi-objective optimization unit, a supplier performance evaluation unit, a supply chain risk early warning unit, and a blockchain traceability platform interface unit. The intelligent material sourcing unit integrates a large-scale pre-trained language model based on Transformer. The multi-objective optimization unit uses the NSGA-III algorithm to generate Pareto optimal solution sets. The supplier performance evaluation unit dynamically assesses supplier quality risks based on Bayesian networks. The supply chain risk early warning unit integrates external data sources and performs risk prediction using Prophet and XGBoost models. The blockchain traceability platform interface unit interacts with blockchain nodes of upstream and downstream enterprises in the supply chain to ensure traceability of material information throughout the entire process. The end-to-end quality closed-loop control module, built upon the PDCA cycle, includes a digital twin modeling unit, a quality planning and management unit, a machine vision online inspection unit, a statistical process control (SPC) unit, a quality improvement traceability unit, and a closed-loop improvement knowledge base. The digital twin modeling unit constructs a digital twin model of the production process using a combination of physical modeling and data-driven approaches, supporting simulation verification of improvement measures. The machine vision online inspection unit integrates high-resolution industrial cameras, multispectral imaging equipment, and a deep learning model based on the visual Transformer. The SPC unit supports real-time calculation of EWMA dynamic quality control charts and process capability indices. The quality improvement traceability unit, by calling the product genealogy intelligent engine module, traces the risks of similar and derivative products when quality problems are discovered. The closed-loop improvement knowledge base records the root cause analysis results and improvement measures for each quality problem, and supports intelligent recommendation of historical solutions through a similarity retrieval algorithm. The low-code configuration and personalized adaptation module allows front-line managers and workshop operators to customize quality reports, monitoring dashboards, alarm rules, and process parameter thresholds through visual drag-and-drop. The system automatically generates the corresponding front-end interface and back-end logic without the need for professional developers. At the same time, it supports dynamically adjusting the complexity of the user interface and the level of detail of the operation instructions according to the roles of different positions and skill levels of personnel, reducing the threshold for front-line employees to use the system and improving system acceptance and usage efficiency.

[0032] Optionally, the data hub and governance module further includes a data lineage tracing unit. This unit constructs a full-link data lineage graph from the original data source to the final data application by parsing ETL data transformation logic, API call chains, and database transaction logs, supporting field-level data traceability. When data anomalies are detected, the lineage graph quickly locates the source and scope of impact of the anomaly and automatically generates a data problem diagnosis report. The data hub and governance module also includes a data asset management unit, which is used to classify and manage various data assets in the data lake, and set differentiated access control policies and data anonymization rules according to data sensitivity and business importance to ensure data security and compliance.

[0033] Optionally, the product genealogy intelligent engine module further includes a design reuse recommendation unit. Based on the feature vector similarity of product nodes, the design reuse recommendation unit automatically retrieves existing design modules in historical products whose similarity to the current design requirements exceeds a threshold during the new product concept design stage. It then pushes reusable design suggestions and historical application cases, including design drawings, process parameters, quality performance, and cost data, to R&D engineers. This supports R&D personnel in assessing the feasibility of reuse and assists in quickly completing derivative product designs through similarity ranking and difference analysis. The design reuse recommendation unit also supports the standardized encapsulation of reusable design modules, forming an enterprise standard parts library and a general module library, continuously enriching the enterprise's intellectual assets.

[0034] Optionally, in the edge-cloud collaborative real-time monitoring module, the streaming computing unit employs a device state recognition algorithm based on a hidden Markov model when calculating the overall equipment efficiency (OEE). This algorithm automatically distinguishes between different states such as equipment operation, standby, idling, shutdown, fault, and replacement, avoiding OEE calculation distortion due to misjudgment of the state. For equipment fault states, the algorithm identifies the causes and frequency patterns of faults through correlation analysis and pushes predictive maintenance suggestions, including recommended maintenance times, spare parts lists, and maintenance operation procedures. The streaming computing unit also supports complex event processing (CEP), defining abnormal event patterns across devices and processes. When an event sequence matching the pattern is detected, it provides early warning of potential production interruption risks.

[0035] Optionally, the full-process quality closed-loop control module further includes a quality cost optimization unit. This unit, based on a dynamic balance model of prevention costs, appraisal costs, and internal and external loss costs, uses a multi-objective optimization algorithm to solve for the optimal quality control level, minimizing the total quality cost while ensuring product quality meets requirements. The quality cost optimization unit comprehensively considers the quality levels of different suppliers, the process capability index of different processes, and the repair costs of different defect types, dynamically adjusting the incoming material inspection sampling plan, process monitoring frequency, and finished product inspection standards to achieve optimal quality cost configuration. The system also includes a lightweight deployment version for SMEs, employing containerized packaging and a microservice architecture. It supports on-demand module loading and cloud-edge collaborative deployment, reducing the technical threshold and initial investment costs for SMEs' digital upgrades, and providing standard data interfaces for rapid integration with industry-standard ERP and inventory management systems.

[0036] Compared with the prior art, the present invention has the following beneficial effects: Compared with the prior art, the present invention has the following beneficial technical effects: This invention constructs a multi-protocol adaptable interface cluster, supporting over ten industrial protocols such as Modbus, OPC UA, and Profinet, enabling unified data acquisition from production equipment of different brands and eras. It specifically addresses the compatibility issues of older equipment by adding multi-functional acquisition terminals. Employing a CDC-based heterogeneous database synchronization mechanism and Merkle tree consistency verification, it ensures eventual consistency of data synchronization between heterogeneous systems such as ERP, PLM, and WMS, reducing data synchronization latency to within seconds. Furthermore, it innovatively introduces a BERT-based semantic similarity matching engine to automatically resolve material coding inconsistencies, achieving a coding mapping accuracy of over 95%, completely breaking down information barriers and forming a unified master data view for the enterprise.

[0037] This project marks the first application of graph neural networks to product derivative relationship mining, constructing a product family tree knowledge graph. Through node embedding representation learning, reusable design modules are automatically identified, increasing design reuse rate by over 30% and significantly shortening new product development cycles. When quality issues arise, graph neural network similarity retrieval allows for the tracing of potential risks across all derivative products within milliseconds, achieving a tracing accuracy rate of 98%, thus transitioning from "passive response" to "proactive early warning." It also supports the assessment of the cascading impact of design changes, preventing systemic quality problems caused by unsynchronized design changes.

[0038] Edge computing nodes were deployed for data preprocessing, employing a dual verification approach using both the Isolation Forest and DBSCAN algorithms to remove outliers, improving data quality by over 80%. A streaming engine based on Apache Flink enabled millisecond-level real-time computation, with key performance indicator (KPI) latency below 100ms. An innovative adaptive threshold optimization mechanism based on a deep Q-network was introduced, reducing false alarm rates by 60% and false negative rates by 45%. Multimodal push notifications ensured that anomaly information reached relevant personnel within 5 seconds, improving production process anomaly response speed by over 5 times.

[0039] Integrating a large-scale pre-trained language model based on Transformer, semantic parsing of material descriptions is performed, achieving a material matching accuracy of over 92%. Employing the NSGA-III multi-objective optimization algorithm, comprehensively considering cost, quality, delivery, and ESG ratings, a Pareto optimal supplier list is generated, reducing overall procurement costs by 15% and improving supplier quality performance by 25%. Integrating external data sources, the Prophet time-series model and XGBoost classification model are used for dynamic assessment and early warning of supplier bankruptcy risk and logistics disruption risk, achieving an average risk identification lead time of 30 days and reducing supply chain disruption events by over 40%. A trusted traceability platform is built based on blockchain technology to ensure the immutability of material information throughout the entire process, meeting the traceability compliance requirements of the high-end manufacturing sector.

[0040] A comprehensive quality management system covering incoming material inspection, process control, and finished product inspection was built based on the PDCA cycle, and a digital twin model of the production process was constructed simultaneously. In the planning phase, LSTM neural networks were used to predict key quality control points, improving the scientific accuracy of control target setting by 35%. In the execution phase, a deep learning model based on visual Transformer was integrated to achieve real-time identification of 19 types of defects, with an accuracy rate of 99.2% and a detection speed of 200 pieces / minute. In the inspection phase, EWMA dynamic quality control charts were used, improving the sensitivity of process anomaly detection by 50%. In the improvement phase, the digital twin model was used for simulation verification of improvement measures, shortening the solution verification cycle from 7 days to 4 hours and increasing the success rate of improvement measures to 85%. A closed-loop improvement knowledge base was established, and historical solutions were intelligently recommended through similarity retrieval, reducing the recurrence rate of similar problems by 70%.

[0041] An innovative low-code configuration platform was introduced, allowing frontline managers to customize quality reports, monitoring dashboards, and alarm rules via drag-and-drop, reducing report generation time from 3 days to 2 hours. The user interface complexity was dynamically adjusted based on different job positions and skill levels, shortening new employee onboarding training time from 2 weeks to 2 days. System usage increased from 45% at the beginning of implementation to over 95%, fundamentally resolving the awkward situation of "system implementation leading to decreased efficiency," ensuring data source accuracy and continuous effective system operation.

[0042] Utilizing containerized packaging and a microservice architecture, it supports on-demand module loading and cloud-edge collaborative deployment, reducing initial investment costs by 70% and shortening the implementation cycle from 6 months to 1 month. It provides standard data interfaces for rapid integration with mainstream ERP and inventory management systems, addressing the technical limitations of SMEs. Built-in industry best practice templates support rapid replication and promotion, and have been successfully implemented in over 20 SMEs across 6 industries, including home appliances, electronics, and auto parts, increasing the digital transformation success rate from the industry average of 35% to 85%.

[0043] In summary, this invention, through the organic integration of innovative technologies such as data governance, product genealogy, edge-cloud collaboration, intelligent supply chain, digital twin quality closed loop, and low-code configuration, comprehensively solves the seven core problems existing in the current MES system. It realizes digital management of the entire manufacturing chain, all elements, and the entire process, significantly improving production efficiency, product quality, supply chain collaboration level, and enterprise core competitiveness. It has important technical value and economic significance for promoting the high-quality development of the manufacturing industry. Attached Figure Description

[0044] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0045] Figure 1 This is a flowchart of a method provided in an embodiment of the present invention.

[0046] Figure 2 This is a system structure diagram provided for an embodiment of the present invention. Detailed Implementation

[0047] 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, and 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.

[0048] The purpose of this invention is to provide a new digital management method and system for manufacturing that can solve several problems existing in current digital management methods and systems for manufacturing.

[0049] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0050] Example 1: This embodiment uses a large electronics manufacturing company (hereinafter referred to as Company A) as an application scenario to explain in detail the specific implementation process and technical effects of the manufacturing digital management method based on the MES system of the present invention. The method flowchart and system structure diagram are as follows: Figure 1 and Figure 2 As shown, Company A primarily manufactures precision structural components for consumer electronics. It possesses a complete production line encompassing injection molding, CNC machining, surface treatment, and assembly, with over 2,000 pieces of equipment including injection molding machines, CNC machining centers, and robots of various brands and eras. It generates approximately 1.5 billion data entries daily. Before implementing this invention, Company A faced severe problems such as data silos, difficulties in quality traceability, and low supply chain collaboration efficiency.

[0051] I. Specific Implementation of Comprehensive Data Governance and Intelligent Integration.

[0052] Company A deployed a data hub and governance module on its production site. For over 2000 production devices, the following steps were implemented: 1. Multi-protocol Adaptation Data Acquisition: For new equipment with digital interfaces (60%), a multi-protocol adaptation interface cluster is deployed to automatically identify the device's communication protocol (Modbus TCP, OPC UA, Profinet, etc.) and establish a standardized data acquisition channel. For older equipment without digital interfaces (40%, mainly injection molding machines and CNC equipment over 10 years old), a multi-functional data acquisition terminal is installed. This terminal integrates a vibration sensor (measurement range 0-20mm / s, accuracy ±0.5%), a temperature sensor (measurement range -20-300℃, accuracy ±0.5℃), and a current transformer (measurement range 0-100A, accuracy ±0.5%). It acquires the device's analog signals at a sampling frequency of 100Hz through an analog-to-digital converter and uses an OPC UA protocol conversion module to convert the acquired data into a standard format and upload it to the data center.

[0053] Results data: Through the above deployment, Company A achieved full data collection from more than 2,000 devices, increasing the device network connectivity rate from 35% before implementation to 98.5%, and the data collection integrity rate reached 99.2%.

[0054] 2. Heterogeneous Database Synchronization: Company A's original ERP system used an Oracle database, its MES system used a MySQL database, and its PLM system used a MongoDB database to store semi-structured data. After implementing this invention, a Debezium-based CDC synchronization mechanism was established to capture change events in each source database in real time. For data synchronization from MongoDB to MySQL, a custom schema mapping strategy was adopted to flatten the nested JSON structure into relational tables. Simultaneously, a Merkle tree-based data consistency verification mechanism was established, performing data reconciliation every 5 minutes and automatically compensating for discrepancies.

[0055] Results: CDC synchronization latency remained stable within 500ms, and data consistency verification pass rate reached 99.97%. In one database migration, the amount of synchronized data reached 2.3TB, and a total of 157 discrepancies were identified and compensated during reconciliation, ensuring cross-system data consistency.

[0056] 3. Material Coding Semantic Matching: Company A has three independent material coding systems: ERP, PLM, and WMS. The codes for the same material are completely different in different systems. After implementing this invention, a semantic similarity matching engine based on a BERT pre-trained model is constructed. Taking the material description text as input, for example, "stainless steel screw M3×8 Phillips head countersunk head" in the ERP system and "Phillips head countersunk head screw M3×8 stainless steel" in the PLM system, the semantic similarity is calculated to be 0.92 (threshold 0.85) using the BERT model, and candidate mapping relationships are automatically generated. After manual confirmation by material engineers, the mapping relationships take effect, and a total of 32,000 material mapping relationships have been established.

[0057] Results data: The material code matching accuracy rate reached 96.5%. After the material master data was unified, the material misissue rate caused by coding errors decreased from 2.3% before implementation to 0.08%, avoiding direct economic losses of approximately 2.4 million yuan per year.

[0058] II. Specific implementation of product genealogy deep mining based on graph neural networks.

[0059] Company A's products exhibit typical characteristics of derivative development. A certain model of mobile phone uses its mid-frame as the base platform (parent product), from which multiple customized versions (child products) are derived for different customers. Each customized version may further generate multiple iterative versions (grandchild products) due to process improvements. The implementation steps are as follows: 1. Product Family Tree Knowledge Graph Construction: Product data from 2018 to 2024 was extracted from the MDM and PLM systems, including product codes, design change records, BOM differences, and process route change records. A product family tree knowledge graph was constructed using the Neo4j graph database, containing 21,000 product nodes and 38,000 derived relationship edges. For example, the basic product P100 serves as the parent node, deriving 12 child nodes such as P110, P120, and P130. Each child node further derives an average of 5.6 grandchild nodes.

[0060] 2. Graph Neural Network Embedding Learning: A graph convolutional network (GCN) is used to learn the embedding representation of the product genealogy knowledge graph. The embedding dimension is set to 128 dimensions, using a 2-layer GCN with ReLU activation, an Adam optimizer, a learning rate of 0.01, and training for 50 epochs. Potential reusable design modules are automatically identified through node similarity calculation. For example, although P100 and P150 have no direct derivation relationship, their feature vector similarity reaches 0.87, and the system automatically indicates a potential opportunity for design module reuse.

[0061] Results data: According to the R&D staff, the system recommended 287 potential reusable modules and confirmed 213 reusable modules, improving the design reuse rate by 28.6% and shortening the average design cycle of new products from 45 days to 32 days.

[0062] 3. Quality Issue Traceability: Company A received a customer complaint that a batch of P123 products had surface scratches. The system automatically extracted the feature vector of product P123 and used a graph neural network model to perform a similarity search in the genealogy knowledge graph, identifying all 23 derivative products that shared the same parent (P100) and the same offspring relationships (P121, P122, P124, P125) as P123. Based on historical quality data analysis, it was predicted that the potential probability of 7 of these products having the same scratch risk exceeded 70%, and differentiated early warning information was pushed. After on-site investigation, it was confirmed that 5 of these products did indeed have the same problem, and corrective measures were taken in a timely manner, avoiding a potential quality loss of approximately 5.8 million yuan.

[0063] III. Specific Implementation of Real-Time Data Processing and Adaptive Monitoring in Edge-Cloud Collaboration: Company A deployed 20 edge computing nodes, each covering 100-150 devices, and deployed a centralized management platform in the cloud.

[0064] 1. Data Preprocessing: Edge nodes perform real-time preprocessing of the collected data. An isolated forest algorithm is used to detect outliers, with a contamination rate parameter set to 0.01 and a window size of 200 sampling points. For identified outliers, a secondary verification is performed using the DBSCAN algorithm (ε=0.3, min_samples=5) to distinguish the anomaly type. For example, an abnormal vibration of a CNC spindle was confirmed through analysis to be a progressive anomaly caused by tool wear, rather than a sensor malfunction.

[0065] Results: After data preprocessing, the amount of data uploaded to the cloud was reduced by 72%, and the accuracy of outlier identification reached 94.3%.

[0066] 2. Adaptive sliding window Cpk calculation: Taking the dimensional parameters of a key process (upper specification limit USL=10.05mm, lower specification limit LSL=9.95mm) as an example, an adaptive sliding window Cpk calculation is implemented. The base window size W_base=50, adjustment coefficient α=0.5, and target process standard deviation σ_target=0.012. The current process standard deviation σ_current=0.018, so the adjusted window size W=50×(1+0.5×0.018 / 0.012)=50×1.75=87.5, rounded to 88. The process mean μ=10.01, σ=0.017, and Cpk=min((10.05-10.01) / (3 0.017),(10.01-9.95) / (3 0.017))=min(0.04 / 0.051,0.06 / 0.051)=min(0.78,1.18)=0.78, which is lower than the preset threshold of 1.0, so an alarm is triggered immediately.

[0067] Performance data: Compared with a fixed window, the adaptive sliding window improves the sensitivity to process fluctuations by 42% and detects process anomalies an average of 23 minutes earlier.

[0068] 3. Threshold Optimization Based on DQN: A deep Q-network model is constructed. The state space includes 18 dimensions such as real-time values, historical mean, and standard deviation of each monitoring indicator. The action space represents the adjustment amount of the threshold for each indicator (discretely divided into 11 levels from -10% to +10%). The reward function is calculated based on the alarm accuracy. After 3 months of online learning (approximately 216,000 samples), the system automatically optimized the alarm thresholds for 136 monitoring indicators.

[0069] Results data: The false alarm rate decreased from 38% before implementation to 12.6%, the false alarm rate decreased from 15% to 6.8%, and the average number of valid alarms per day decreased from 235 to 82, significantly reducing the burden on maintenance personnel.

[0070] IV. Specific Implementation of Intelligent Supply Chain Collaboration Based on Large Model and Multi-Objective Optimization

[0071] 1. Intelligent Material Sourcing: Company A processes approximately 12,000 purchase orders monthly, involving over 6,800 material categories. After implementing this invention, a large-scale pre-trained language model based on Transformer (350 million parameters) was integrated and fine-tuned for training on material description texts in the manufacturing industry (training data of 200,000 annotated corpora). Taking the material description "304 stainless steel precision shaft, diameter 5±0.02, length 45±0.1" in a purchase requisition as an example, the system automatically parses out: material category "precision shaft," material "304 stainless steel," key parameters "diameter 5mm tolerance ±0.02mm, length 45mm tolerance ±0.1mm," and identifies synonyms such as "stainless steel shaft" and "shaft parts."

[0072] Results data: The accuracy rate of automatic material classification reached 96.2%, the accuracy rate of key parameter extraction was 91.5%, and the processing time for sourcing a single material was reduced from an average of 8 minutes to 2.3 seconds.

[0073] 2. Multi-objective Optimization of Supplier Selection: For a specific batch of precision shaft procurement, six qualified suppliers participated in the bidding. The system collected historical supplier data: quality performance (batch pass rate over the past 12 months), on-time delivery rate, price, and ESG rating. The NSGA-III algorithm was used for multi-objective optimization, with a population size of 100, 200 generations, a crossover probability of 0.9, and a mutation probability of 0.1. In the Pareto optimal solution set generated, supplier B (cost 28 yuan / piece, quality risk 0.12, on-time delivery rate 97%) and supplier D (cost 31 yuan / piece, quality risk 0.08, on-time delivery rate 99%) were selected as the preferred options. The procurement engineer selected supplier D based on the current priority objective (quality is the primary consideration for this batch).

[0074] Results data: After implementing NSGA-III multi-objective optimization, Company A's overall supplier satisfaction (weighted by cost, quality, and delivery) increased by 18.6%, and quality risk decreased by an average of 23.5%.

[0075] 3. Supply Chain Risk Early Warning: The system integrates external data sources to obtain daily information such as raw material price indices, logistics freight rate indices, and weather warnings. The system detected continuous heavy rain warnings in the region where a key supplier is located (weather risk probability 85%), and the supplier's historical delivery data showed a 32% increase in delivery delay rate during the rainy season. The system comprehensively calculates the risk level: β=5, P=0.85, S=0.7, AlertLevel=5×0.85×0.7=2.975=Level 3 (preset threshold Level 3). The system automatically triggers the emergency plan: recommends 3 alternative suppliers, suggests adjusting the safety stock of the material from 7 days to 15 days, and initiates a rapid certification process for alternative materials.

[0076] Results data: This early warning was issued 14 days in advance, helping Company A to complete emergency material preparation before the logistics were disrupted by the rainstorm, thus avoiding a production loss of approximately 18 million yuan.

[0077] V. Specific implementation of full-process quality closed-loop control based on PDCA and digital twin.

[0078] 1. Digital Twin Modeling: Taking a certain injection molding production line of Company A as an example, a digital twin model based on a combination of physical modeling and data-driven approach is constructed. The physical modeling uses Moldex3D software to simulate the injection molding process, while the data-driven model uses an LSTM neural network (15-dimensional input, 128-dimensional hidden layers, and 5-dimensional output), and integrates actual production data (mold temperature, pressure, holding time, cooling time, etc.) for model correction.

[0079] Results data: The digital twin model has a prediction error of less than 0.03mm for product size and an accuracy of 89.5% for predicting the probability of defect occurrence.

[0080] 2. PDCA closed-loop control: Planning Phase P: The system uses an LSTM model (24 months of historical data, 30-day time window) to predict the control target for a key quality control point of a certain product model—flatness (maximum specification 0.15mm). Based on the prediction results, the flatness control target for the current month is set as a mean of 0.08mm and a standard deviation of 0.015mm.

[0081] Phase D: Deploy a machine vision online inspection system on the production line, integrating five high-resolution industrial cameras (5 megapixels) and multispectral imaging equipment (including visible, infrared, and ultraviolet bands). Employ an attention-based visual Transformer model (trained on 200,000 labeled defect images, covering 18 defect categories) to acquire product images in real time for defect identification. The inspection speed reaches 210 pieces per minute, with a defect identification accuracy of 98.7%, including a 96.2% identification rate for scratches and a 99.1% identification rate for dirt / stains.

[0082] Inspection Phase C: The system automatically summarizes quality data and generates an EWMA dynamic quality control chart. On a certain day, it is found that the flatness EWMA statistic increases for five consecutive points. Although it does not exceed the control limit, the trend is abnormal. The system automatically issues a yellow warning.

[0083] Phase A of Improvement: Through product family tree knowledge graph tracing, it was discovered that the problematic product shared a parent-generation relationship with another customer complaint from the previous week regarding excessive flatness. Root cause analysis simulation using a digital twin model confirmed that the problem stemmed from a deviation in mold temperature setting (actual 235℃ vs. standard 240℃). The system automatically triggered a process adjustment procedure, correcting the mold temperature to 240℃, and the improvement effect was verified. The average flatness decreased from 0.093mm to 0.081mm, and the standard deviation decreased from 0.021mm to 0.016mm. The improvement measures and their effects were recorded and entered into the closed-loop improvement knowledge base.

[0084] Results data: After implementing PDCA closed-loop control, the product defect rate of this production line decreased from 1.8% to 0.63%, the cost of quality loss decreased by 64%, and the recurrence rate of similar problems decreased by 81%.

[0085] VI. Specific implementation of low-code configuration and personalized adaptation.

[0086] Company A's workshop has 426 operators and 58 managers, with significant differences in job skills. After implementing the low-code configuration platform: Team leader A customized the team quality dashboard by dragging and dropping, selecting to display six indicators such as the shift's OEE, the top 3 defects, and the current quantity of work-in-process. The configuration took 25 minutes.

[0087] On the third day of new employee B's employment, the system automatically simplified the interface based on his job position (CNC operator) and skill level (beginner), displaying only necessary operation instructions and 5 key monitoring indicators, along with illustrated operation steps. Employee B commented, "The system is easier to use than I expected; I don't need to remember so much."

[0088] Senior employee C is highly experienced, so the system automatically activates the advanced mode, displaying more process parameters and analysis charts. Wang can quickly query historical process records through the system.

[0089] Results data: System usage increased from 42% at the beginning of implementation to 97.3%, the average onboarding time for new employees was shortened from 11 days to 3.5 days, and employee satisfaction rating increased from 3.2 out of 5 to 4.7.

[0090] VII. Specific Implementation of Lightweight Deployment for SMEs: Company B, a small and medium-sized enterprise (80 employees, annual output value of approximately 50 million yuan), adopted the lightweight deployment version of this invention. It uses containerized packaging (Docker + Kubernetes) and loads modules on demand: first, the data acquisition and real-time monitoring module was deployed (2 weeks), followed by the gradual loading of the quality management module (1 week) and the supply chain collaboration module (1 week). Cloud-edge collaborative deployment: two servers were deployed on the edge side (total investment of 48,000 yuan), and public cloud resources were rented in the cloud (average monthly cost of 1,200 yuan). A standard API interface was provided for integration with Company B's existing Kingdee KIS inventory management system, and data integration was completed in 2 days.

[0091] Results data: Company B's initial investment in digital upgrade was 53,000 yuan (only 15% of the traditional MES solution), with an implementation period of 1 month (the traditional solution requires 4-6 months). Six months after implementation, the overall equipment efficiency increased by 18.3%, the defect rate decreased by 1.2 percentage points, and the inventory turnover rate increased by 22.6%.

[0092] In summary, this embodiment verifies the feasibility and significant technical effects of the present invention through specific data, comprehensively solves the seven core problems raised in the background technology, and realizes digital management of the entire manufacturing chain, all elements, and the entire process.

[0093] The foregoing has provided a detailed description of a manufacturing digital management method and system based on a MES system provided by this invention. Specific examples have been used to illustrate the principles and implementation methods of this application. The descriptions of the above embodiments are only for the purpose of helping to understand the method and its core ideas; furthermore, those skilled in the art will recognize that, based on the ideas of this application, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of this application.

[0094] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the descriptions are relatively simple; relevant parts can be referred to the method section.

[0095] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the method and core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A digital management method for manufacturing based on a MES system, characterized in that, Includes the following steps: Comprehensive Data Governance and Intelligent Integration: A data hub encompassing multiple heterogeneous data sources is constructed, including at least production equipment, Enterprise Resource Planning (ERP), Product Lifecycle Management (PLM), Warehouse Management System (WMS), and Supply Chain Management (SCM). A multi-protocol compatible interface cluster is deployed to collect heterogeneous data from production equipment of different brands and eras in real time. For older equipment lacking digital interfaces, multi-functional data acquisition terminals integrating vibration, temperature, and current sensors are installed. Analog signals from the equipment are acquired through analog-to-digital conversion, and the collected data is converted to a standard format using an OPC UA protocol conversion module. A heterogeneous database synchronization mechanism based on Change Data Capture (CDC) is established, capturing and writing data to the target database in real time when changes occur in the source database. A Merkle Tree-based data consistency verification mechanism is also established, ensuring eventual consistency of cross-system data synchronization through periodic reconciliation and automatic compensation for discrepancies. For the issue of inconsistent material coding between different systems, a semantic similarity matching engine based on a BERT pre-trained model is constructed to automatically calculate the semantic similarity of material descriptions in different coding systems, generate candidate mapping relationships, and implement them after manual confirmation, forming a unified master data view. Product family tree deep mining based on graph neural network: Extract product design change history, BOM material bill difference and process route change records from the master data management system (MDM) and product lifecycle management system (PLM), and construct a product family tree knowledge graph using the graph database Neo4J. The product family tree knowledge graph is composed of product nodes and edges representing derivative relationships. Among them, the new generation of products developed based on a brand-new design concept is used as the parent node, the second generation of products formed by partial design changes based on the first generation of products is used as the child node, and the third generation of products formed by further design changes based on the second generation of product nodes is used as the grandchild node. A graph neural network (GNN) is used to embed the product family tree knowledge graph into a representation learning process to generate a feature vector for each product node. By calculating node similarity, potential reusable design modules and common problems across generations are automatically identified. Each product node is associated with its full lifecycle quality traceability data, process documents, change history, and customer complaint records to construct a holographic profile of the product. Real-time data processing and adaptive monitoring through edge-cloud collaboration: Edge computing nodes are deployed on the production site to preprocess collected equipment status data, process parameters, and quality inspection data. This includes outlier removal based on the isolated forest algorithm, multi-source timestamp calibration based on the NTP protocol, and data compression based on differential coding. Sliding window aggregation analysis is performed using an Apache Flink-based streaming computing engine to calculate in real-time the overall equipment efficiency (OEE), the process capability index (Cpk) of key process parameters, and the defect rate. The formula for calculating the process capability index (Cpk) is as follows: Where USL is the upper specification limit, LSL is the lower specification limit, μ is the process mean, and σ is the process mean; the process mean and process standard are calculated based on real-time sampled data within an adaptive sliding window, and the window size adjustment formula is: Where W is the adjusted sliding window size, W base The base window size, α is the adjustment factor, and σ is the base window size. current Let σ be the current process standard deviation. target The target process standard deviation is set; when the Cpk value is detected to be lower than the preset threshold or the quality defect rate exceeds the control limit, an alarm is immediately triggered and sent to the relevant workstation terminals, management personnel's mobile terminals and workshop dashboards via multimodal push; Intelligent supply chain collaboration based on large-scale models and multi-objective optimization: Upon receiving a production order, the Bill of Materials (BOM) is automatically parsed. A large-scale pre-trained language model based on the Transformer architecture is integrated to standardize the semantics of material descriptions, enabling automatic material classification, key parameter extraction, and synonym recognition. The system then uses supplier product catalog data from a digital supplier management platform for intelligent matching and sourcing. Combining supplier historical quality performance KPIs, on-time delivery rates, price factors, and ESG (Environmental, Social, and Governance) ratings, a non-dominated sorting genetic algorithm (NSGA-III) with an elitist strategy is employed for multi-objective optimization. The optimization objectives are to minimize procurement cost (C), minimize quality risk (R), and maximize on-time delivery rate (D). This generates a Pareto optimal solution set as a preferred supplier list. The quality risk (R) is dynamically assessed based on supplier historical quality performance data using a Bayesian network model. The risk assessment formula is as follows: in, P ( F i Let be the probability of the i-th type of quality failure event. I i The impact of the i-th type of quality failure event is determined; production plans, material demand forecasts and inventory status are shared in real time through the supply chain collaboration platform, and purchase order generation and payment processes are automatically triggered based on smart contract technology to drive suppliers to carry out JIT just-in-time delivery; End-to-end quality closed-loop control based on PDCA and digital twin: based on the international quality management standard ISO The 9001 methodology and the PDCA cycle are used to construct a full-chain digital quality management system covering supplier incoming material inspection, production process quality control, and finished product outgoing inspection, while simultaneously building a digital twin model of the production process. In the planning phase (P), historical quality data is used to predict key quality control points and set dynamic control targets via a Long Short-Term Memory (LSTM) network. In the execution phase (D), a machine vision online inspection system integrating high-resolution industrial cameras and multispectral imaging equipment acquires product images in real time, and uses an attention-based visual Transformer model for defect identification, supporting real-time detection and classification of various defect types such as scratches, dirt, damage, burrs, color differences, and dimensional deviations. In the inspection phase (C), quality data is automatically summarized and a dynamic quality control chart based on the Exponentially Weighted Moving Average (EWMA) is generated to monitor process stability in real time. In the improvement phase (A), when a systematic quality deviation is detected, the potential risks of similar and derivative products are traced through the product family tree knowledge graph and graph neural network model. Root cause analysis and improvement measure simulation verification are performed using the digital twin model, automatically triggering design changes or process adjustments, and recording the improvement measures and their effects to form a closed-loop improvement knowledge base.

2. The manufacturing digital management method based on MES system according to claim 1, characterized in that, The comprehensive data governance and intelligent integration further includes: establishing a data quality management system, setting data quality dimensions including completeness, accuracy, consistency, timeliness, and uniqueness; scanning data in the data lake through scheduled tasks, marking, alerting, or automatically repairing data that does not conform to the rules based on preset rules; for outliers in the collected data, in addition to using the isolated forest algorithm, a density-based clustering algorithm DBSCAN is also used for secondary verification, and the identified outliers are traced back to their source, distinguishing three different causes: sensor failure, communication interruption, and production process abnormality, and triggering equipment maintenance work orders, network diagnostic processes, or quality warnings respectively; and establishing a data lineage tracing mechanism, generating a data lineage map by analyzing the data flow path and transformation logic, supporting end-to-end traceability from the final report to the original data source, ensuring data interpretability and efficient problem location.

3. The manufacturing digital management method based on MES system according to claim 1, characterized in that, The in-depth product genealogy mining based on graph neural networks further includes: when product quality issues or customer complaints occur, the system automatically extracts the feature vector of the problematic product, performs similarity retrieval in the product genealogy knowledge graph through a graph neural network model, identifies all derivative products with the same parent, child, or grandparent relationship as the problematic product, and predicts their potential risk probability based on historical quality data, pushing differentiated early warning information; the early warning information includes risk level, possible failure modes, suggested stricter sampling inspection schemes, and solutions to similar historical issues; at the same time, the system identifies the chain reaction of design changes through correlation analysis. When the design of a parent product changes, the system automatically assesses the impact on all child and grandchild products, pushes design coordination notifications, and avoids quality problems caused by design changes not being synchronized to derivative products.

4. The manufacturing digital management method based on MES system according to claim 1, characterized in that, The edge-cloud collaborative real-time data processing and adaptive monitoring further includes: a dynamic optimization mechanism for monitoring thresholds based on reinforcement learning, which learns the correlation between historical alarm data and actual faults through a deep Q-network (DQN) model, automatically optimizing the alarm thresholds of each monitoring indicator to minimize the false alarm rate while ensuring no missed alarms; when a continuous downward trend in Cpk or periodic fluctuations in the quality defect rate are detected, potential root causes are identified through a time-series causal inference algorithm, including factors such as raw material batch fluctuations, changes in environmental temperature and humidity, wear and tear of equipment tools, and changes in operators, and pushed to process engineers for confirmation; the multimodal push methods include push notifications from industrial tablets at workstations, vibration alerts from wearable devices, audible and visual alarms in the workshop, and mobile APP notifications, automatically selecting the optimal push channel based on the alarm level and personnel role.

5. The manufacturing digital management method based on MES system according to claim 1, characterized in that, The intelligent supply chain collaboration based on large models and multi-objective optimization further includes: integrating external data sources, including macroeconomic data, raw material price indices, logistics freight rate indices, meteorological data, and geopolitical risk indices; and using the time series forecasting model Prophet and the machine learning classification model XGBoost to dynamically assess and warn of supplier bankruptcy risk, raw material shortage risk, logistics disruption risk, and delivery delay risk. The risk warning level is calculated comprehensively based on the probability of occurrence and the degree of impact, using the following formula: Among them, AlertLevel is the risk warning level, with a value of an integer from 1 to 5, β is the scaling factor, P is the probability of risk occurrence, and S is the degree of risk impact. When the risk level exceeds the preset threshold, the emergency plan is automatically triggered, including recommending alternative suppliers, adjusting safety stock levels, initiating alternative material certification processes, or changing production plans. At the same time, a reliable traceability platform for the supply chain is built based on blockchain technology to record information on the entire process of materials from raw material procurement, production and processing, quality testing to logistics and distribution, ensuring the immutability and traceability of supply chain data.

6. A manufacturing digital management system based on a MES system, characterized in that, include: The data hub and governance module is used to perform full-domain data governance and intelligent integration, including a multi-protocol adaptation interface cluster, a distributed message queue unit based on Apache Kafka, a unified data lake storage unit, a heterogeneous database synchronization unit, and a data quality governance unit; the multi-protocol adaptation interface cluster supports Modbus and OPC. The system integrates UA, Profinet, EtherCAT, CANopen, and MQTT industrial protocols and JDBC / ODBC database interfaces to achieve interconnectivity with equipment from different brands and heterogeneous information systems. The heterogeneous database synchronization unit employs a CDC-based synchronization mechanism and integrates a Merkle Tree-based data consistency verification engine. The data quality governance unit has a built-in data quality management rule base, supporting the configuration and execution of rules for integrity, accuracy, consistency, timeliness, and uniqueness. It also includes a BERT-based semantic similarity matching engine to address material coding inconsistencies. The unified data lake storage unit integrates a relational database, a time-series database (InfluxDB), a graph database (Neo4j), and a document database (MongoDB) to store structured business data, equipment time-series data, product genealogy knowledge graphs, and unstructured documents, respectively. The product genealogy intelligent engine module interfaces with the Master Data Management System (MDM) and the Product Lifecycle Management System (PLM) to extract product design change history and BOM differences, construct and visualize the product genealogy knowledge graph, including a graph database storage unit, a graph neural network model training unit, and a similarity retrieval unit; the graph neural network model training unit uses a graph convolutional network (GCN) or a graph attention network (GAT) to learn embedded representations of product nodes and their derived relationships. The similarity retrieval unit supports quickly retrieving related derivative products based on feature vector similarity from any product node, and displays their quality traceability data, process documents and change history in association. The edge-cloud collaborative real-time monitoring module is deployed on the edge computing node cluster and cloud centralized management platform at the production site, including a data preprocessing unit, a streaming computing unit, an adaptive monitoring threshold optimization unit, and a multimodal alarm push unit. The data preprocessing unit integrates the Isolation Forest algorithm and the DBSCAN algorithm for outlier detection and removal; the streaming computing unit is based on the Apache Flink engine, supports sliding window aggregation analysis and complex event processing (CEP), and calculates the overall equipment efficiency (OEE), process capability index (Cpk), and quality defect rate in real time; the adaptive monitoring threshold optimization unit dynamically optimizes alarm thresholds based on the Deep Q-Network (DQN) model. The multimodal alarm push unit supports integrated push notifications from workstation industrial tablets, wearable devices, workshop sound and light alarms, and mobile apps. The intelligent supply chain collaboration module interfaces with Supplier Management System (SRM), Enterprise Resource Planning (ERP), and Supply Chain Finance System. It includes an intelligent material sourcing unit, a multi-objective optimization unit, a supplier performance evaluation unit, a supply chain risk early warning unit, and a blockchain traceability platform interface unit. The intelligent material sourcing unit integrates a large-scale pre-trained language model based on Transformer. The multi-objective optimization unit uses the NSGA-III algorithm to generate Pareto optimal solution sets. The supplier performance evaluation unit dynamically assesses supplier quality risks based on Bayesian networks. The supply chain risk early warning unit integrates external data sources and performs risk prediction using the Prophet and XGBoost models. The blockchain traceability platform interface unit is used to interact with blockchain nodes of upstream and downstream enterprises in the supply chain to ensure that the information of materials throughout the entire process is traceable. The end-to-end quality closed-loop control module, built upon the PDCA cycle, includes a digital twin modeling unit, a quality planning and management unit, a machine vision online inspection unit, a statistical process control (SPC) unit, a quality improvement traceability unit, and a closed-loop improvement knowledge base. The digital twin modeling unit constructs a digital twin model of the production process using a combination of physical modeling and data-driven approaches, supporting simulation verification of improvement measures. The machine vision online inspection unit integrates high-resolution industrial cameras, multispectral imaging equipment, and a deep learning model based on the visual Transformer. The SPC unit supports real-time calculation of EWMA dynamic quality control charts and process capability indices. The quality improvement traceability unit, by calling the product genealogy intelligent engine module, traces the risks of similar and derivative products when quality problems are discovered. The closed-loop improvement knowledge base records the root cause analysis results and improvement measures for each quality problem, and supports intelligent recommendation of historical solutions through a similarity retrieval algorithm. The low-code configuration and personalized adaptation module allows front-line managers and workshop operators to customize quality reports, monitoring dashboards, alarm rules, and process parameter thresholds through visual drag-and-drop. The system automatically generates the corresponding front-end interface and back-end logic without the need for professional developers. At the same time, it supports dynamically adjusting the complexity of the user interface and the level of detail of the operation instructions according to the roles of different positions and skill levels of personnel, reducing the threshold for front-line employees to use the system and improving system acceptance and usage efficiency.

7. The manufacturing digital management system based on the MES system according to claim 6, characterized in that, The data hub and governance module further includes a data lineage tracing unit. This unit constructs a full-link data lineage graph from the original data source to the final data application by analyzing ETL data transformation logic, API call chains, and database transaction logs, supporting field-level data traceability. When data anomalies are detected, the lineage graph quickly locates the source and scope of impact of the anomaly and automatically generates a data problem diagnosis report. The data hub and governance module also includes a data asset management unit, which is used to classify and manage various data assets in the data lake, and set differentiated access control policies and data anonymization rules according to data sensitivity and business importance to ensure data security and compliance.

8. The manufacturing digital management system based on the MES system according to claim 6, characterized in that, The product genealogy intelligent engine module further includes a design reuse recommendation unit. Based on the feature vector similarity of product nodes, the design reuse recommendation unit automatically retrieves existing design modules in historical products whose similarity to the current design requirements exceeds a threshold during the new product concept design stage. It then pushes reusable design suggestions and historical application cases to R&D engineers, including design drawings, process parameters, quality performance, and cost data. This supports R&D personnel in assessing the feasibility of reuse and assists in quickly completing derivative product designs through similarity ranking and difference analysis. The design reuse recommendation unit also supports the standardized encapsulation of reusable design modules, forming an enterprise standard parts library and a general module library, continuously enriching the enterprise's intellectual assets.

9. The manufacturing digital management system based on the MES system according to claim 6, characterized in that, In the edge-cloud collaborative real-time monitoring module, the streaming computing unit uses a device state recognition algorithm based on a hidden Markov model when calculating the overall equipment efficiency (OEE). This algorithm automatically distinguishes different states such as equipment operation, standby, idling, shutdown, fault, and replacement, avoiding OEE calculation distortion due to misjudgment of the state. For equipment fault states, the algorithm identifies the causes and frequency patterns of the fault through correlation analysis and pushes predictive maintenance suggestions, including recommended maintenance time, spare parts list, and maintenance operation procedures. The streaming computing unit also supports Complex Event Processing (CEP), which defines abnormal event patterns across devices and processes. When an event sequence that matches the pattern is detected, it provides early warning of potential production interruption risks.

10. The manufacturing digital management system based on the MES system according to claim 6, characterized in that, The full-process quality closed-loop control module further includes a quality cost optimization unit. This unit, based on a dynamic balance model of prevention costs, appraisal costs, and internal and external loss costs, uses a multi-objective optimization algorithm to solve for the optimal quality control level, minimizing the total quality cost while ensuring product quality meets requirements. The quality cost optimization unit comprehensively considers the quality levels of different suppliers, the process capability index of different processes, and the repair costs of different defect types, dynamically adjusting the incoming material inspection sampling plan, process monitoring frequency, and finished product inspection standards to achieve optimal quality cost configuration. The system also includes a lightweight deployment version for SMEs, employing containerized packaging and a microservice architecture. It supports on-demand module loading and cloud-edge collaborative deployment, reducing the technical threshold and initial investment costs for SMEs' digital upgrades, and providing standard data interfaces for rapid integration with industry-standard ERP and inventory management systems.

Images