An aviation manufacturing industry known source brain data intelligent manufacturing management method

Abstract

This invention relates to the field of knowledge-based brain data management technology, and discloses a knowledge-based brain data-driven intelligent manufacturing management method for the aviation manufacturing industry. It utilizes natural language processing technology to analyze internal data, constructs a large-scale knowledge base for the aviation manufacturing field, and builds various training algorithm models for the aviation manufacturing field on this basis. The purpose of this invention is to address the problems of knowledge acquisition, knowledge integration, and knowledge application in the aviation manufacturing field, as well as the personalized needs of aviation users, by providing a thinking mode that conforms to human users. This method can reduce production consumption, improve product testing efficiency, product qualification rate, production yield, quality inspection efficiency, and testing accuracy, and achieve a more advanced and efficient intelligent manufacturing management method for knowledge-based brain data in the aviation manufacturing industry.

Description

Technical Field

[0001] This invention relates to the field of knowledge-based brain data platform technology, specifically, to a knowledge-based brain data-driven intelligent manufacturing management method for the aerospace manufacturing industry. Background Technology

[0002] The organic coupling and integration of technologies such as the Internet of Things (IoT), big data, artificial intelligence (AI), and the industrial internet with digital twin-related technologies such as manufacturing system modeling, simulation, virtual reality (VR), augmented reality (AR), and intelligent control makes it possible to establish parallel-operating manufacturing digital twin systems in virtual space. With the development of next-generation information technologies such as cloud computing, IoT, big data, and AI, the connotation of intelligent manufacturing has undergone a rapid transformation. Intelligent manufacturing refers to a production process that integrates knowledge engineering, manufacturing software systems, and robot vision technologies, and is independently realized by intelligent robots without human intervention. As one of the most important technologies in intelligent manufacturing, an intelligent manufacturing system is a system that integrates intelligent machines and intelligent behavior, and can be integrated into all aspects of the manufacturing process, such as order processing, product design, production, marketing, and sales, to fully leverage the manufacturing capabilities of advanced production systems in a flexible manner. The foundation of intelligent manufacturing is the digitization of problems in industrial production, using the acquired data to model solutions, and implementing application services such as graph retrieval and document retrieval based on domain knowledge graphs and document libraries. The data here refers to all quantifiable indicators used in production, such as equipment instructions, machining process parameters, and executable decisions. Quantitative data derived from knowledge is used to correct and improve the manufacturing process, fundamentally preventing problems from occurring.

[0003] The impact of big data technology on the aviation manufacturing industry is mainly twofold: first, it improves product quality and production efficiency through the application of big data technology in manufacturing and management processes; second, it integrates big data technology into existing products and services for disruptive innovation. In short, intelligent manufacturing incorporating big data technology focuses on two key aspects: manufacturing process technology and the technology of the manufactured products. Currently, scientific data services face challenges such as weak central system construction, low levels of unified management, non-standardized services, fragmented data resources, inconsistent data quality, weak global influence, reliance on others for core technologies, and a lack of high-level, multi-skilled talent. Aviation and machinery manufacturing, facing the field of industrial big data, urgently need to utilize mature artificial intelligence, big data, data standardization, and data mining technologies to create a scientific data knowledge service platform that can strengthen and standardize scientific data management, ensure scientific data security, improve openness and sharing, and better support technological innovation in the aviation manufacturing sector. Products in aviation, machinery, and other industrial manufacturing sectors involve high-precision, high-reliability products across multiple technological fields, characterized by complex structures, long production cycles, and multiple production states. With the development of the information age, the volume of quality information data generated during the production and research process is becoming increasingly massive. This data includes key information such as the causes of various quality problems, problematic components, and measures taken. Due to the current lack of unified data management, various types of quality information are scattered across business systems, existing in the form of electronic or paper documents. It is extremely difficult to control the complex process of integrating multi-source heterogeneous data, standardizing quality control management, customizing multi-level data review, and managing the entire lifecycle of data resources, from data collection to quality control, review, release, issuance, and service.

[0004] The Industrial Brain is an integrated computing platform based on Alibaba Cloud's big data. Through a data factory, it aggregates data from various industrial enterprises, including enterprise system data, factory equipment data, sensor data, and personnel management data. Leveraging voice interaction, image / video recognition, machine learning, and artificial intelligence algorithms, it unlocks the value of massive amounts of data. Currently, although data is being innovatively used in many areas, such as transmitting collected production parameters to the Industrial Brain, its overall utilization rate remains relatively low, leaving much of its economic value untapped. The modern aircraft manufacturing process is essentially a process of digital product modeling, data transmission, extension, and processing; the final aircraft product can be seen as the material manifestation of data. With the application of PDM technology, manufacturing process information and processes are managed under the PDM system. The PadMan system transforms information or data according to a specific process, reflecting the essence of the modern aircraft manufacturing process. The PadMan system is a multi-user, network-based management system primarily for project management. The PadMan information management system utilizes the Single Source of Product Data (SSPD) concept to build its database, thus forming a relatively complete manufacturing process information generation, process flow design, and process data management system in the aircraft manufacturing field. The development of the PadMan system was undertaken in a completely new technological field, involving a wide range of aspects, significant technical challenges, and a heavy workload. The process information primarily focuses on aircraft assembly process information, and has not yet been extended to the management of parts process information, resulting in significant waste of financial and human resources and extended manufacturing cycles. Because the data from aircraft design and manufacturing has not been effectively planned and integrated into a unified database, some issues remain unresolved in the software. For example, the management of data related to parts manufacturing and tooling manufacturing is not yet included in the system; change management only records and tracks changes, and the control over changes is insufficient and has not extended to deeper or more granular levels. Furthermore, integration with ERP software is needed to complete other aspects of manufacturing information management.

[0005] Aviation equipment manufacturing is a complex systems engineering project involving numerous stages, high resource requirements, and significant challenges in process control. Due to its cross-business, multi-entity, multi-level, and strongly coupled characteristics, it is susceptible to various uncertainties and disturbances, exhibiting significant dynamic and nonlinear features. This leads to multi-level and multi-dimensional complexity issues in the modeling, optimization, and closed-loop collaborative control of complex manufacturing systems. Current production management methods, primarily based on human intervention, lack overall system control, often resulting in insufficient resource support and delays in the timely arrival of various manufacturing resources, hindering smooth production and limiting workshop efficiency. The long development cycles, cross-business domains (design, process, production, quality, testing, maintenance, etc.), multiple stakeholders (factories, institutes, suppliers, etc.), and stringent quality control of aviation products and their manufacturing processes result in complexity and numerous abnormal disturbances. Traditional manufacturing models, relying primarily on human experience for decision-making and control, are no longer adequate for the current intelligent transformation and development of aviation manufacturing. Furthermore, the high precision requirements and difficulty in accurately controlling deformation during the machining of large and complex aviation structural components further complicate the process. In traditional manufacturing processes, the processing, monitoring, and optimization control stages are independent of each other, making it impossible to achieve real-time control, optimization, and adjustment of the manufacturing process, and thus difficult to guarantee the machining accuracy of parts. In particular, the numerous electromechanical components in aerospace manufacturing, the high intensity of work, and the increasing service life lead to frequent failures, and the causes of failure are diverse. In many cases, the lack of professional personnel to assist in the analysis makes it difficult to find the true cause of the failure.

[0006] With the widespread application of modern computer technology in aircraft manufacturing, the correct storage, maintenance, and accurate and timely transmission of data have become crucial issues that modern aircraft manufacturing must address. This necessitates the establishment of an effective system for managing assembly data in modern aircraft manufacturing. Currently, my country has not yet researched the unified management of product data in the aircraft manufacturing sector. To address the core issues of intelligent manufacturing in industry, a data intelligence product, the Zhiyuan Knowledge Brain, has been developed to accelerate the construction of new industrial infrastructure. The Zhiyuan Knowledge Brain mainly comprises two parts: the Knowledge Brain Data Processing and Management Platform and the Knowledge Brain Intelligent Application Platform. The Knowledge Brain Intelligent Application Platform is a technology for implementing knowledge application systems based on knowledge engineering in aerospace manufacturing enterprises. The aircraft manufacturing information management system serves every process node of the enterprise's operation through process service interfaces, knowledge-based job accompaniment, and intelligent knowledge applications, ultimately aiming to improve employee work efficiency. The Brain Data Processing and Management Platform utilizes the artificial intelligence technology of the industrial brain. It requires the input of thousands of production parameters collected in real-time from the workshop into the industrial brain to efficiently and cost-effectively complete image quality inspection, monitor and control variables in real-time during production, and improve production quality and yield. Currently, collecting and using data for a single purpose without sharing it with others for reuse is a waste of resources.

[0007] Aviation manufacturing enterprises are knowledge-intensive, and knowledge is a crucial implicit asset. While aviation manufacturing plants have relatively high levels of automation, they are undergoing a transformation from "manufacturing" to "intelligent manufacturing." This involves intelligent hardware, intelligent management systems, high-speed communication networks, and intelligent data processing and decision support systems. The entire manufacturing process is undergoing automation, digitalization, informatization, and intelligentization, moving towards higher efficiency and quality. Due to the involvement of numerous manufacturers and communication protocols, a large amount of data is accumulated and continuously increasing. The more times data is reused, the greater the risk of misuse. The modern aircraft manufacturing process is essentially a process of digital product modeling, data transmission, extension, and processing. The final aircraft product can be seen as the material manifestation of data. The corresponding work, the required data, and the data generated during the process are processed through networks and server-side software to generate charts. The software's functions are implemented using COM technology. Currently, the output of information data resources is enormous, diverse, scattered, and lacks unified management principles and methods. Product testing, inspection, and debugging have always been bottleneck processes in production. Testing / adjustment costs account for over 30% of total production costs, with an average processing time of 0.5-1 hour per product. These factors increase the difficulty of data collection, resulting in low utilization and hindering transparent factory management. In particular, factors such as different batches of electromechanical components in aerospace manufacturing can cause fluctuations in indicators. Therefore, the large amount of data generated during production must adhere to GMP and safety production management standards, and product development relies heavily on engineers' experience. However, there has been no effective linkage between R&D and manufacturing data, leading to difficulties in process control and hindering the inheritance and stability of subsequent R&D experience. While artificial intelligence algorithms can perform deep learning calculations on all related parameters to accurately analyze the key parameters most relevant to production quality and build parameter curve models to precisely implement optimal parameters in large-scale production, the return on investment for implementing a "bottom-up, modular industrial internet platform" architecture in the daily production processes of industrial manufacturing management—both production line and engineering operations—is not necessarily a definite yes. Taking the "bottom-up" approach to building industrial internet platforms, with the lowest-level OT systems (such as MES and APS) and control systems (such as PLC and SCADA) as an example, data collection from equipment, production lines, and factories is achieved, along with information-based control of factory equipment. However, this approach has also led to a large number of "dumb devices" that cannot collect data, and the collected data existing in different types of equipment or systems. This results in serious "data silos" between different systems, and the implementation cycle is long, highly customized, and expensive, making it unaffordable for most enterprises. A major reason for this situation is that the "bottom-up" approach to building industrial internet platforms lacks top-level design and value-oriented thinking.Integrating core technologies such as big data with intelligent manufacturing to further improve production capacity and quality while reducing costs is a key task for innovation in the aerospace manufacturing industry. Complex analytical applications or decision support typically require highly reliable aerospace manufacturing knowledge graphs. Intelligent question-and-answer applications need to integrate general knowledge graphs, common-sense knowledge graphs, and domain-specific knowledge graphs. The lifecycle of a knowledge graph has clearly defined stages, yet these stages are also interconnected and organically integrated. Therefore, it is necessary to comprehensively analyze and formulate relevant standards, specifications, operating procedures, calling interfaces, data samples, model algorithms, evaluation indicators, and visualization tools from multiple dimensions to ensure that the knowledge graph standard solution possesses high reliability, high reusability, rapid migration, and continuous evolution, optimization, and accumulation capabilities for different business or task scenarios in the aerospace field. Image classification, which distinguishes different categories of images based on their semantic information, is a fundamental problem in computer vision. It involves inspecting product appearance through images to identify defective products. In industrial manufacturing processes, automated optical inspection (AOI) equipment using deep learning technology is beginning to replace traditional inspection equipment. Because manufacturing enterprises typically rely on visual inspection for product quality control during production, the process is costly and inefficient. Statistics from actual manufacturing operations show that over 20% of equipment downtime is caused by excessive tool wear. Natural language itself is highly ambiguous, especially for frequently occurring entities, which may correspond to multiple names, and each name may correspond to multiple entities with the same name. Connecting highly ambiguous natural language with knowledge graphs is a crucial step in constructing and completing knowledge graphs, and the key to achieving this is entity linking technology. Entity linking is a matching and disambiguation task. The resource scheduling and optimization decision-making problem in aerospace manufacturing systems is a complex system optimization and control problem, and its complexity is manifested in…

[0008] 1) The system has a complex composition. The system includes both hardware entities such as control equipment, processing equipment, logistics equipment, warehousing equipment, and auxiliary equipment, and software entities such as production processes and production management standards;

[0009] 2) Complex organizational structure. Hierarchical structure is an essential attribute of discrete manufacturing systems. The system includes various levels such as individual equipment, production units, production lines, production workshops, and supply chains, resulting in complex production and logistics organizational relationships.

[0010] 3) Complex operating mechanism. Under the production model where mass production and R&D coexist, the system is required to have higher production flexibility and reconfigurability to adapt to diverse and customized user needs.

[0011] How to deeply integrate DT (Data Technology) and AI to drive intelligent management and control of manufacturing systems through "dynamic perception, real-time analysis, autonomous decision-making, and precise execution" still presents many key challenges. On the one hand, the difficulty in acquiring, modeling, fusing, and organizing multi-source, multi-dimensional, and heterogeneous data across the entire manufacturing system and its processes hinders the application of AI-based methods due to a lack of data support. On the other hand, the high requirements for reliability, safety, and stability in industrial settings place higher demands on the interpretability and reliability of AI results, limiting the application of related technologies. Aerospace manufacturing systems exhibit typical characteristics of complex mega-systems, with numerous constituent elements and complex relationships, exhibiting typical uncertainty and emergent features. How to establish a reasonable abstract model to realize its abstract expression in virtual space and reflect its operational laws and characteristics remains a problem to be solved. Summary of the Invention

[0012] The purpose of this invention is to provide a data-driven intelligent manufacturing management method for the aerospace manufacturing industry. This invention addresses the issues of knowledge acquisition, integration, and application in the aerospace manufacturing field, as well as the personalized needs of aerospace users. It provides a user-friendly thinking mode that can reduce production consumption, improve product testing efficiency, product qualification rate, production yield, quality inspection efficiency, and testing accuracy, and achieve a more advanced and efficient intelligent manufacturing management platform (MES) for the aerospace manufacturing industry.

[0013] The above-mentioned objective of this invention can be achieved through the following measures: a data-driven intelligent manufacturing management method for the aerospace manufacturing industry, comprising: using natural language processing technology to analyze internal data, constructing a large-scale knowledge base in the aerospace manufacturing field, and building a pre-trained model in the aerospace manufacturing field based on this knowledge base; characterized in that: the aerospace manufacturing field knowledge base effectively plans and incorporates data from aircraft design and manufacturing into a unified database, incorporates data from parts manufacturing, tooling manufacturing, etc., into a tracking and management system, integrates general knowledge graphs, common sense knowledge graphs, and domain knowledge graphs, integrates with ERP software, and completes graph retrieval and document retrieval applications based on the domain knowledge graph and document library. The service manages manufacturing information and processes both structured and unstructured data to form a domain knowledge graph and a domain document library. The pre-trained model uses an intelligent manufacturing management platform to monitor and control received and processed data variables in real time during production, efficiently and cost-effectively completing image quality inspection. Based on the labeled model, it provides information extraction service interfaces, enabling dynamic training and deployment of labeled data. Real-time collected production parameters are transmitted to the industrial brain, where artificial intelligence algorithms perform deep learning calculations on all related parameters, accurately analyzing the key parameters most relevant to production quality. Parameter curve models and algorithm models are built, using test / inspection data as the main body and leveraging IoT and algorithm model technology. The technology performs real-time monitoring and detects key indicators. Using AI image processing, it extracts effective pixels and uploads a domain knowledge graph containing product defects to a cloud computing platform. Algorithm training is then conducted using deep learning and image processing techniques. Data from key production stages is aggregated and integrated via the cloud. Combining the strong fitting capabilities of convolutional neural networks and anomaly detection algorithms, deep learning is performed on massive datasets. Key features are extracted from the original images through convolution operations. Max pooling is then used after these convolution operations, treating the image as a matrix and performing matrix multiplication. After completing the convolution and pooling operations, these images are input into three fully connected layers for two full-connection operations, using different convolution kernels. A convolutional neural network (CNN) is trained on a GPU, and all elements in the matrix are summed to obtain the final result. At the same time, deep computation and analysis are performed on various types of product data to build a fault detection and perception prediction model. This model obtains the demand perception of manufactured products, identifies minor faults in manufactured products, predicts measurement point faults and defects, outputs prediction results, and provides early warnings of faults and defects. Based on the industrial brain, an algorithm optimization model is built. The algorithm model is used to comprehensively analyze and evaluate process capabilities and optimize production processes, turning data into knowledge and knowledge back into data. This accurately and in real time predicts the amount of product defects and equipment failures, makes defect judgments, issues instructions to capture defective products, and provides process parameters to guide the optimal solution for actual production.

[0014] This invention is achieved through the following technical solution: a data-driven intelligent manufacturing management method for the aerospace manufacturing industry, comprising:

[0015] This paper utilizes natural language processing (NLP) technology to analyze internal data and construct a large-scale knowledge base for the aerospace manufacturing field. Based on this, a pre-trained model for the aerospace manufacturing field is built. The key features include: effective planning of aircraft design and manufacturing data within a unified database; inclusion of data on parts manufacturing and tooling manufacturing into a tracking and management system; integration of general knowledge graphs, common-sense knowledge graphs, and domain knowledge graphs; and integration with ERP software to achieve manufacturing information management based on domain knowledge graphs and document libraries, enabling graph retrieval and document retrieval application services. Simultaneously, it processes structured and unstructured data to form domain knowledge... Image recognition and domain documentation libraries are provided. Pre-trained models utilize the Industrial Brain platform to monitor and control received and processed data variables in real time during production, efficiently and cost-effectively completing image quality inspection. Based on the labeled model, information extraction service interfaces are provided, enabling dynamic training and deployment of labeled data. Real-time collected production parameters are transmitted to the Industrial Brain, where artificial intelligence algorithms perform deep learning calculations on all related parameters, accurately analyzing the key parameters most relevant to production quality. Parameter curve models and algorithm models are built, using test / inspection data as the main body, leveraging IoT and algorithm model technology for real-time monitoring and detection of indicators. AI image technology is used to improve... Valid pixels are extracted, and the domain knowledge graph containing product defects is uploaded to a cloud computing platform. Algorithm training is performed using deep learning and image processing technologies. Data from key production stages is aggregated and integrated via the cloud. Combining the strong fitting ability and anomaly detection algorithms of convolutional neural networks, deep learning is applied to massive datasets. Key features are extracted from the original images through convolution operations. Max pooling is then used after the convolution operations, treating the image as a matrix and performing matrix multiplication. After completing the convolution and pooling operations, these images are input into three fully connected layers for two full-connection operations. The convolutional neural network is trained on a GPU using different convolutional kernels. The system uses a CNN to accumulate all elements in the matrix to obtain the final result. Simultaneously, it performs deep computation and analysis on various product data to build a fault detection and perception prediction model. This model acquires demand perception of manufactured products, identifies minor faults, predicts measurement point faults and defects, outputs prediction results, and provides early warnings of faults and defects. Based on an industrial brain, an algorithm optimization model is built. This model is used for comprehensive analysis and evaluation of process capabilities and optimization of production processes. Data is transformed into knowledge, and knowledge is then transformed back into data. This allows for accurate and real-time prediction of product defect quantities and equipment failures, defect determination, issuing instructions to capture defective products, and providing optimal solutions for actual production based on process parameters.

[0016] To better realize this invention, the cloud computing platform, based on the convolutional neural network algorithm for natural image recognition, divides the data to be trained into equal parts, corresponding one-to-one with the nodes on the cloud computing platform, and stores them in an evenly distributed manner. The training of the network is completed by using the data of the CNN network stored on each node. After the task receives the data, the operation module manages the large dataset of the tree diagram structure through a master node. The master node distributes the operation tasks to each sub-node. After the sub-nodes complete the data processing, they are aggregated and sent back to the master node for processing. The processed task is decomposed into multiple task modules and used on the nodes on the platform. The local changes in weights and biases are calculated through forward and backward propagation to form the value of the intermediate key. After all samples have been calculated, local file processing is performed. The processed local file summarizes the data obtained from each training session and writes it into a global file.

[0017] To better realize this invention, the product data, originally physically distributed across multiple databases, is further organized into a logically strictly constrained single database, resulting in a logically single product data source (SSPD). The SSPD serves as the core access source for the underlying data and related product data of the entire aerospace manufacturing product system. At the outset of creating the knowledge graph, a knowledge system is created, with each project corresponding to a unique knowledge system. This knowledge system acts as the trunk of the knowledge graph, while each file and its text paragraphs are considered its leaves. "Categories" and "Attributes" are dragged into the operation area, and names and types are defined for "Entity Categories" and "Attributes." Connections are established from entities to attributes, defining entity-related attributes and establishing relationships. Connections are also established from one entity to another, and relationship names are defined in the file template panel. The knowledge system is named by selecting the method for building the knowledge system, and the data form is imported. Provide a sample file template. Create relationship category files and attribute category files according to the format of the sample file. In the knowledge system list, locate the knowledge system to be created, download the data files used to create the knowledge system, select "File Category," "Document Creation Time," "Knowledge System Tags," "Report Type," and whether "OCR Recognition" is required, and upload the knowledge system attribute category files and knowledge system relationship category files respectively. Create the knowledge system by uploading files and dragging and dropping. The completed knowledge system will be applied to the project as the knowledge system of the project's knowledge graph. It will store product-related data between product data distributed in different databases, forming multiple distributed databases, and building a system database that runs through all data sources in the entire process of aircraft manufacturing data management. This ensures that the product data is complete, consistent, and reliable, meeting the needs of aviation enterprises from customer selection to aircraft delivery and service support.

[0018] To better implement this invention, furthermore, when uploading a file, in the file management list, select the file to be started with OCR, select "Open OCR Recognition," enter the editing state, and after editing, click any area outside the input box to end the editing state. The system constructs a multi-dimensional feature vector based on network condition metrics, normalizes each dimension, calculates the mean and variance of each dimension in the sequence used for training, and then performs convolution operations using one-dimensional causal convolution and dilated convolution. After fusing the various channels, the uploaded data image file is obtained, which is then automatically converted into editable text on the computer. After uploading the file, the recognition status of the current document can be viewed in the document list.

[0019] To better realize this invention, furthermore, when creating the knowledge graph, entities, attributes, and relationships are obtained from the document through annotation tools, and data is added by importing triples; click Project Management - Import Triples to upload the organized triple data, and generate a complete knowledge graph based on the triples; document annotation is divided into two steps: first, entity and attribute annotation is completed, and then relationship annotation is performed. After both steps are completed, an annotation task is completed; the entity annotation method is as follows: first, select a word segment in the entity annotation area, select an entity category for the word segment, and then generate an entity card in the card display area. Below the generated entity card is the attribute card, where the corresponding attribute name is selected and attribute value is created; select the attribute name, which is the attribute name defined in the knowledge system that is associated with the entity; the system automatically associates the annotation information with the knowledge system according to the current annotation status of the project, forming the domain knowledge graph of the project.

[0020] To better realize this invention, the domain knowledge graph further obtains entity names through the NER model. The NER model first determines the required entity name "mention," defines the digital information of the geometry, topology, materials, processes, and relevant technical specifications of the manufactured product, and stores them in the Engineering Data Set and the Automatic Parts List (APL), respectively. Then, each "mention" is matched to its corresponding candidate entity in the knowledge graph, sorted, aggregated, and fused with multi-dimensional and heterogeneous manufacturing process data, and sequence labeling modeled on a character-by-character basis. The pre-trained model receives training data in the form of text data and dictionary tuples, updates the existing model according to the context to train NER, and loops on the training data. A smart progress table (tqdm) is added before the data loader of the iterator to obtain sufficient iterations. The tqdm is applied to deep learning, and an iterative progress bar (tqdm()) function is used to create a progress bar to save the training process information.

[0021] To better realize this invention, the pre-trained model processes the maximum activation feature map of the convolutional layer during training, then occludes the input image. The occluded new image is then used as the new input to the network to continue training the model. In the image recognition process, to further accelerate the training speed, the image is preprocessed to remove redundant and interference information. Then, the processed image is divided into recognition regions. The state at the next moment is predicted using the latest historical data, and the prediction error is calculated. The historical sequence data from time t-k+1 to time t is normalized, and an unstructured text data output sequence is obtained based on the normalization. Professional information entities of the processing platform architecture are extracted from the unstructured text data, and these entities are stored in a relational database.

[0022] To better implement this invention, the pre-trained model is further based on a character-level Chinese dataset. It uses a trained recognizer to perform entity recognition tasks, predicts entity strings, and obtains a list of predicted entity sets, [setpre1, setpre2, ..., setpreren]. The real entity strings are extracted to obtain a list of real entity sets, where setprei represents the unique set of all entities extracted from a sample.

[0023] To better realize this invention, the perception and prediction model further collects multi-dimensional dynamic data throughout the entire process. Through parallel mapping of the virtual space of the physical manufacturing system, it performs deep knowledge mining to predict future data changes. Then, a trained recognizer compares the predicted data with future real data to determine if anomalies occur. Based on the data predicted by the prediction module at time t+1 and the corresponding error matrix, it determines whether the data acquired at time t+1 is abnormal. The covariance matrix is ​​calculated based on the error matrix, and different prediction and discrimination models are learned for different network environments. Multiple indicators are used to measure the network from different perspectives, and a multivariate Gaussian distribution is used to establish the relationship between these indicators, accurately detecting network anomalies and automatically adapting to different network environments to achieve proactive perception of the manufacturing system. This is combined with deep learning frameworks for learning. By analyzing the changing patterns of data, the system aggregates and loads dynamic and static full-volume, multi-dimensional, and multi-scale information from the manufacturing system to form control strategies. These strategies rely on automated rule-based events for processing, providing real-time feedback and real-time sensing data control, simulation data, and multiple data feedback mechanisms for manufacturing site reflections, major equipment failure shutdown control, missing parts alarms, broken tool alarms and decisions, and safety alarms and decisions. This constructs a multi-dimensional fusion intelligent agent component encompassing the physical, information, and business spaces of the manufacturing system. From the spatiotemporal domains, it builds intelligent agent model components for manufacturing resources, manufacturing units, and the supply chain within the intelligent manufacturing space. Around the dynamic evolution of logistics, value stream, information flow, and business flow within the manufacturing system, it establishes a dynamic collaborative operation mechanism centered on intelligent agents, enabling real-time synchronous simulation and virtual-physical linkage control and information interaction transmission mechanisms for multiple elements, businesses, and processes within the intelligent manufacturing space.

[0024] To better realize this invention, the algorithm optimization model, driven by the AI ​​brain of the manufacturing system, synchronously collects information across the entire domain and runs intelligent agent control commands around the workshop logistics, information flow, and business flow. Through closed-loop simulation decision-making and heterogeneous integration between the manufacturing system and the AI ​​brain, the optimized control commands are transmitted to the intelligent agent model components of the manufacturing unit and supply chain, driving the operation of the physical manufacturing system. Through virtual-real fusion, the simulation and optimization decision-making and adaptive adjustment model of the manufacturing system are driven to achieve adaptive optimization of process parameters. The optimized parameters are fed back to the processing equipment through CNC commands, realizing closed-loop control of the entire processing process. The fault detection module uses knowledge graphs to expand the description of fault events and faulty components. It integrates and filters data existing in various fault diagnosis reports and fault case libraries, extracts useful information, and organizes it as the basis for knowledge construction. The knowledge base is used to more accurately locate the system failure parts and their causes. A fault library is built by uploading files. The fault detection module compares the information in the fault library with the information in the fault library to determine whether it is a repeated fault. The fault information includes: product type, product name, discovery time, fault description, estimated loss, and preliminary handling status.

[0025] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0026] (1) This invention delves into the aviation field, transforming massive amounts of raw data rich in unit information into a "entity-attribute-relationship" graph data form containing higher-dimensional information through steps such as entity recognition, relation extraction, entity fusion, and attribute fusion. This weaves the raw data into a knowledge graph with industry attributes, thereby enabling a qualitative leap in the performance and utility boundaries of basic functions such as data fusion, information retrieval, interactive analysis, and multi-dimensional display, ultimately achieving ultra-deep analysis and application of industry data. By exploring the interrelationships between multiple metrics through multivariate Gaussian distribution, the accuracy of detection is improved. Furthermore, the learning and detection process can be completed automatically online, improving the classification performance of various convolutional neural network models on different datasets. The trained model also exhibits excellent robustness in recognizing randomly occluded images and can adapt to various network environments.

[0027] (2) This invention addresses the issues of knowledge acquisition, knowledge integration, and knowledge application in the field of aviation manufacturing. It delves into the aviation field and uses a knowledge base in the aviation manufacturing field to effectively plan the data in aircraft design and manufacturing and incorporate it into a unified database. It also incorporates data on parts manufacturing, tooling manufacturing, etc. into a tracking and management system. It integrates general knowledge graphs and common sense knowledge graphs into domain knowledge graphs and ERP software. It uses COM technology to control these software programs in the program, so that they can generate the charts that users want according to the data in the database, thereby expanding the functions of the software itself and making the generation and browsing of charts more convenient.

[0028] (3) This invention achieves manufacturing information management by completing graph retrieval and document retrieval application services based on domain knowledge graphs and document libraries. At the same time, it processes structured and unstructured data to form domain knowledge graphs and domain document libraries, thereby realizing ultra-deep analysis and application of industry data. Through entity recognition, relation extraction, entity fusion, and attribute fusion, it transforms the data into a "entity-attribute-relationship" graph data form containing higher-dimensional information. Its essence is to weave the original data into a knowledge graph with industry attributes, thereby enabling a qualitative leap in the performance and utility boundaries of basic functions such as data fusion, information retrieval, interactive analysis, and multi-dimensional display. It integrates general knowledge graphs, common sense knowledge graphs, and domain knowledge graphs with intelligent behavior, conforms to the thinking mode of human users, and has a powerful ability to integrate multi-source heterogeneous data. Through in-depth big data mining and a dual quality control approach combining automation and manual methods, standardized storage, unified display, and open sharing of product demand perception across different types of data resources are achieved. This allows for flexible and full utilization of the manufacturing capabilities of advanced production systems, integrating across all aspects of the manufacturing process, including order processing, product design, production, marketing, and sales. Quantitative data derived from knowledge is used to correct and improve the manufacturing process, fundamentally preventing problems from arising. Project management manages all data and documents throughout the knowledge graph's lifecycle. In project management, a knowledge graph related to knowledge and documents can be constructed by creating a project, uploading a knowledge system, and selecting associated documents. The knowledge graph can expand the descriptions of fault events and faulty components, integrating and filtering data from various fault diagnosis reports and fault case libraries, extracting useful information, and organizing it as the foundation for knowledge construction. This knowledge base is used to more accurately locate system failure points and their causes. The integration of language knowledge graphs, factual knowledge graphs, domain knowledge graphs, and newly acquired knowledge from data sources is required at different granularities depending on the application scenario. Meanwhile, as the underlying support platform for knowledge graph solutions, the knowledge graph storage and computing tools offer rich business functions and components, perfectly combining business knowledge with machine intelligence. Based on business scenarios and data graph characteristics, they provide powerful functions such as relationship network analysis, real-time multi-dimensional retrieval, and information comparison and collision, making the rapid implementation of knowledge graph industry solutions possible. Furthermore, based on big data analysis, they can achieve full-process lifecycle management of the aerospace manufacturing industry from "demand to manufacturing to demand," enabling numerous core intelligent manufacturing applications such as digital, networked, and intelligent production optimization, precise matching, supply chain optimization, and marketing push. They provide end-users with personalized services such as subscriptions, proactive push notifications, thematic databases, personal knowledge management, and personal work centers, thereby driving the design and development of aerospace manufacturing products. Simultaneously, in the unified data storage process, the concept of a single product data source is used to ensure data storage consistency and reduce data redundancy to a certain extent.Meanwhile, the use of ODBC technology in the client-database communication process simplifies the portability of application software to other database systems. This significantly improves software maintainability, reusability, and scalability, leaving ample room for program upgrades and improvements. Furthermore, it enhances the visualization of the application interface, making it easier to operate. The graphical user interface allows users to easily complete the design and management process and quickly generate desired graphs and tables. This promotes the digital engineering transformation of the aviation industry, enabling deep cloudification of aviation enterprise equipment, production lines, and all aspects of R&D and manufacturing, truly realizing the digital transformation and upgrading of aviation enterprises and solving problems related to innovation, speed, efficiency, quality, and cost in production and operation.

[0029] (4) This invention combines knowledge and experience in the aviation field, and improves the knowledge system and domain knowledge expression system in terms of concept identification, domain applicability, level of detail, and system construction, in accordance with aviation data standards. Using an intelligent manufacturing management platform, the received and processed data variables are monitored and controlled in real time during the production process, image quality inspection is completed efficiently and at low cost, production quality is improved, and information extraction related service interfaces are provided based on the annotation model to realize dynamic training and deployment of annotation data, and finally the optimal parameters can be accurately implemented in large-scale production. The production parameters collected in real time are transmitted to the industrial brain, and the artificial intelligence technology of the industrial brain is used to improve the productivity of aviation manufacturing products. Through artificial intelligence algorithms, deep learning calculations are performed on all related parameters to accurately analyze the key parameters most related to production quality, and parameter curve models and algorithm models are built. Variables are monitored and controlled in real time during the production process. Data linkage is achieved through the integration between various systems, forming a complete intelligent manufacturing implementation solution for the process industry. A large number of artificial intelligence algorithms can be used to classify, cluster, associate, and predict data. With the accumulated basic data, in-depth mining and utilization can be carried out to achieve the purpose of predicting production and sales data, analyzing quality problems, and optimizing product formulas. Knowledge fusion requires support for result traceability and data confidence analysis. A combination of model evaluation algorithms and human feedback mechanisms ensures quality control and version evolution optimization of the knowledge graph. When manufacturers conduct product development and process improvement, they can use the system's predictive simulation capabilities to extract, classify, and model product data, predicting the suitability of the current development or process optimization. Simultaneously, the system provides suggested optimization parameters to help manufacturers improve expected development values, complete process optimization, or product development, reducing resource waste from multiple development attempts. It can also reuse and integrate data from public and private sector channels and employ modern analytical techniques. Merging multiple datasets allows for more accurate and frequent coverage of a wider population. Leveraging these data synergies can generate tangible benefits.

[0030] (5) This invention takes test / detection data as the main body, uses IoT and algorithm model technology for real-time monitoring and detection of indicators, uses AI image technology to extract effective pixels, and changes different convolution kernels to ensure that all the necessary features are extracted, which can reduce a lot of computation and improve accuracy. Moreover, it can determine what number the image represents without having to know the situation of all points, which can further improve the running speed and accuracy. The domain knowledge graph with product defects is uploaded to the cloud computing platform, and the algorithm is trained through deep learning and image processing technology. Data from key production links is aggregated and connected through the cloud. Combining the strong fitting and anomaly detection algorithm of convolutional neural network, deep learning is performed on massive data. At the same time, deep calculation and analysis are performed on various types of product data to build a fault detection and perception prediction model. This application of convolution operation to neural network is the convolutional neural network, which is equivalent to turning multiple data into one data, greatly reducing the amount of data. Pooling can reduce the amount of data by 75%. Convolutional neural network has a strong ability to extract features, and features are easy to be extracted repeatedly. Therefore, it is equivalent to the neural network being trained repeatedly, and the neural network has a good prediction effect on the training data. By acquiring demand perception of manufactured products, minor faults can be identified, and defects at measurement points can be predicted and warned. The accuracy of identification can reach over 95%, reducing the defect rate by 30%-50%. This real-time monitoring and detection of indicators based on big data and artificial intelligence algorithm models allows for accurate prediction. Combining the strong fitting ability of convolutional neural networks and the strong generalization ability of anomaly detection algorithms, it predicts and analyzes large amounts of data. Compared to previous predictions using data statistics and traditional machine learning, the detection process requires no manual intervention, and the average number of detected indicators is significantly reduced, greatly lowering the maintenance cost per major event. This increases manufacturing efficiency and reduces production costs. Through deep learning and image processing technology for algorithm training and cloud computing, the average pass rate is improved by 5-8 percentage points. Deep learning enables faster network convergence and higher prediction accuracy and robustness. Furthermore, the overall debugging efficiency of aerospace manufacturing products is optimized by 30%-38%.

[0031] (6) This invention aims to achieve standardized knowledge management and operation in the field of industrial big data. Based on the industrial brain, it constructs an algorithm optimization model to promote the online platform operation of resources such as knowledge, products, and services. The algorithm model is used to conduct comprehensive analysis and evaluation of process capabilities and optimize production processes, turning data into knowledge and knowledge back into data. It accurately and in real time predicts product defect quantity and equipment failure, makes defect judgments, issues instructions to capture defective products, and provides optimal solutions for actual production by providing process parameters. This method digitizes problems in aerospace production and uses the acquired data to model solutions to problems. When similar problems occur, it can propose solutions based on the model. In this way, knowledge is solidified in data and algorithm models. Past experience is transformed into exploitable and quantifiable continuous value, so that knowledge is not lost due to human factors. Then, data is transformed into knowledge, extending from "solving obvious problems" to "exploring invisible, hidden problems." Finally, knowledge is transformed back into data. Using a network architecture combining client / server and web methods, quantitative data derived from knowledge is used to correct and improve the manufacturing process, managing aircraft manufacturing processes. This ensures that data generation and management are standardized, with a large amount of computation handled by servers and clients, fully leveraging the efficiency of each node in the network. This fundamentally avoids problems and reduces time wasted due to repeated coordination between different process stages. This revolutionizes the traditional design-manufacturing-testing model for equipment, shifting towards a new model of design-virtual integration-digital manufacturing-physical manufacturing. Furthermore, knowledge is accumulated and refined within the artificial intelligence model, solving the long-standing problem of knowledge accumulation and optimization in aviation industry production. This improves the reliability of artificial intelligence results and significantly reduces the interpretability requirements of artificial intelligence models in aviation industry applications. Attached Figure Description

[0032] The present invention will be further described in conjunction with the following drawings and embodiments. All inventive concepts of the present invention should be considered as disclosed content and within the scope of protection of the present invention.

[0033] Figure 1 This is a schematic diagram of the architecture of the Zhiyuan Brain Data Intelligent Manufacturing Management Platform for the aviation manufacturing industry of this invention.

[0034] Figure 2 It is based on the functional flowchart of the Zhiyuan Knowledge Brain Data Processing Platform. Detailed Implementation

[0035] To more clearly illustrate the technical solutions of the embodiments of the present invention, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. It should be understood that the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments, and therefore should not be regarded as a limitation on the scope of protection. 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.

[0036] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "set up," "connected," and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0037] Example 1:

[0038] This embodiment presents a data-driven intelligent manufacturing management method for the aerospace manufacturing industry. It utilizes natural language processing technology to analyze internal data, constructing a large-scale knowledge base for the aerospace manufacturing field. Based on this, a pre-trained model for the aerospace manufacturing field is built. The aerospace manufacturing knowledge base effectively plans and incorporates data from aircraft design and manufacturing into a unified database. Data on parts manufacturing and tooling manufacturing are included in a tracking and management system. It integrates general knowledge graphs, common-sense knowledge graphs, and domain knowledge graphs, and integrates with ERP software to achieve manufacturing information management based on domain knowledge graphs and document libraries, enabling graph retrieval and document retrieval application services. Simultaneously, it processes structured and unstructured data to form a domain knowledge graph and... Domain documentation library; pre-trained models utilize an intelligent manufacturing management platform to monitor and control received and processed data variables in real time during production, efficiently and cost-effectively completing image quality inspection. Based on the labeled model, it provides information extraction service interfaces, enabling dynamic training and deployment of labeled data. Real-time collected production parameters are transmitted to the industrial brain, where artificial intelligence algorithms perform deep learning calculations on all related parameters, accurately analyzing the key parameters most relevant to production quality. Parameter curve models and algorithm models are built, using test / inspection data as the main body, leveraging IoT and algorithm model technology for real-time monitoring and detection of indicators. AI image technology is used to extract... Effective pixels are used to upload a domain knowledge graph containing product defects to a cloud computing platform. Algorithms are trained using deep learning and image processing technologies. Data from key production stages is aggregated and integrated in the cloud. Combining the strong fitting ability and anomaly detection algorithms of convolutional neural networks, deep learning is performed on massive datasets. Key features are extracted from the original images through convolution operations. Max pooling is then used after the convolution operations. During convolution, the image is treated as a matrix, and matrix multiplication is performed. After completing the convolution and pooling operations, these images are input into three fully connected layers for two full-connection operations. Different convolutional kernels are used to train the convolutional neural network on a GPU. The CNN then sums all elements in the matrix to obtain the final result. Simultaneously, it performs deep computation and analysis on various product data to build a fault detection and perception prediction model. This model acquires the demand perception of manufactured products, identifies minor faults in manufactured products, predicts measurement point faults and defects, outputs prediction results, and provides early warnings of faults and defects. Based on the industrial brain, an algorithm optimization model is built. This model is used to comprehensively analyze and evaluate process capabilities and optimize production processes, turning data into knowledge and knowledge back into data. It accurately and in real time predicts product defect quantity and equipment failures, makes defect judgments, issues instructions to capture defective products, and provides optimal solutions for actual production by providing process parameters.

[0039] Example 2:

[0040] This embodiment further optimizes upon embodiment 1. The cloud computing platform uses a convolutional neural network algorithm for natural image recognition. The data to be trained is divided into equal parts, corresponding one-to-one with nodes on the cloud computing platform, and stored in an evenly distributed manner. The training of the network is completed using the data of the CNN network stored on each node. After the task receives the data, the operation module manages the large dataset of the tree diagram structure through a master node. The master node distributes the operation tasks to each sub-node. After the sub-nodes complete the data processing, they are aggregated and sent back to the master node for processing. The processed task is decomposed into multiple task modules and used on the nodes on the platform. The local changes in weights and biases are calculated through forward and backward propagation to form the value of the intermediate key. After all samples have been calculated, local file processing is performed. The processed local file summarizes the data obtained from each training session and writes it into a global file.

[0041] The other parts of this embodiment are the same as those in Embodiment 1, so they will not be described again.

[0042] Example 3:

[0043] The knowledge base in the aerospace manufacturing field stores complete product information in the Single Product Data Source (SSPD), and different supporting data for the product can be extracted from different usage perspectives. The Single Product Data Source (SSPD) is not a data source stored in a separate database, but rather a logically unified data source.

[0044] This embodiment further optimizes upon Embodiment 1 or 2 described above. The aerospace manufacturing knowledge base uses the logically single product data source SSPD as the core access source for the underlying data and related product data of the entire aerospace manufacturing product system. Specifically, in this embodiment, product data that was originally physically distributed across multiple databases is organized into a logically strictly constrained single database, and a knowledge system is created at the beginning of the knowledge graph creation. Each project corresponds to a unique knowledge system. The knowledge system is used as the trunk of a knowledge graph, with each file and paragraph of text within a file serving as its leaves. "Categories" and "Attributes" are dragged and dropped into the operation area. Names and types are defined for "Entity Categories" and "Attributes." Connections are made from entities to attributes, defining entity-related attributes and connecting entities from one entity to another. Relationship names are defined in the file template panel. The knowledge system is named by selecting a construction method. Data forms and sample file templates are imported, and a sample file is provided. Relationship category files and attribute category files are created according to the sample file's format. The knowledge system is located in the list, and the data files used for its creation are downloaded. "File Category," "Document Creation Time," "Knowledge System Tags," "Report Type," and whether "OCR Recognition" is required are selected. The knowledge system attribute category file and relationship category file are uploaded. The knowledge system is created through file upload and drag-and-drop. The completed knowledge system will be applied to the project, serving as the knowledge system of the project's knowledge graph. Product-related data is stored across product data distributed in different databases, forming multiple distributed databases and constructing a database that runs throughout the entire aircraft manufacturing data management system. This ensures that the product data is complete, consistent, and reliable, meeting the needs of aviation companies in all aspects, from customer selection to aircraft delivery and service support.

[0045] The other parts of this embodiment are the same as those in Embodiment 1 or 2 above, so they will not be described again.

[0046] Example 4:

[0047] This embodiment further optimizes any one of the above embodiments 1-3. When uploading a file, select the file to be started with OCR in the file management list, select to open OCR recognition, enter the editing state, and click any area outside the input box to end the editing state. The system forms a multi-dimensional feature vector based on the network condition measurement index, normalizes each dimension, and calculates the mean and variance of each dimension in the sequence used for training. Then, it performs convolution operations using one-dimensional causal convolution and dilated convolution. After the convolution operation is performed in the convolutional layer, the uploaded data image file is obtained after the various channels are fused. It is then automatically converted into text that can be edited on the computer. After uploading the file, the recognition status of the current document can be viewed in the document list.

[0048] The other parts of this embodiment are the same as any one of the embodiments 1-3 above, so they will not be described again.

[0049] Example 5:

[0050] This embodiment further optimizes any one of the above embodiments 1-4. When creating the map, a complete map based on triplet data is obtained through a labeling tool, and data is added by importing triplet data.

[0051] Specifically, this involves clicking "Project Management" - "Import Triples," uploading the prepared triplet data, and then generating a complete knowledge graph based on that triplet data. Document annotation is divided into two steps: first, complete the entity and attribute annotations, and then annotate the relationships. Once both steps are completed, an annotation task is finished. The entity annotation method is as follows: first, select a word segmentation in the entity annotation area, then select the entity category after segmentation, and then generate a corresponding entity card for that entity in the card display area. Below the generated entity card are attribute cards, where you select the corresponding attribute name and create attribute values. The attribute name is a predefined attribute name associated with the entity in the knowledge system. The system automatically associates the annotation information with the knowledge system based on the current annotation status of the project, forming the domain knowledge graph for that project.

[0052] The domain knowledge graph is processed by a Named Entity Recognition (NER) model based on a pre-trained language model to obtain entity names. The NER model first determines the required entity names (mentions), defining the digital information of the manufactured product's geometry, topology, materials, processes, and relevant technical specifications, storing these in the Engineering Data Set and the Automatic Parts List (APL), respectively. Then, each mention is matched to its corresponding candidate entities in the knowledge graph, sorted, aggregated, and integrated with multi-dimensional, heterogeneous manufacturing process data. Sequence labeling is performed on a character-by-character basis for modeling. The pre-trained model receives training data in the form of text data and dictionary tuples, updating the existing model based on context to train NER. The model iterates through the training data, adding a smart progress meter (tqdm) before the dataloader to obtain sufficient iterations. Tqdm is then applied to deep learning, using the iterative progress bar (tqdm()) function to create a progress bar and save training process information.

[0053] The other parts of this embodiment are the same as any one of the embodiments 1-4 above, so they will not be described again.

[0054] Example 6:

[0055] This embodiment further optimizes any one of embodiments 1-5 above. During the training process, the pre-trained model processes the maximum activation feature map of the convolutional layer and then occludes the input image. The occluded new image is then used as the new input of the network to continue training the model. In the image recognition process, in order to further accelerate the training speed, the image is preprocessed to remove redundant and interference information. Then, the processed image is divided into recognition regions. The state at the next moment is predicted using the latest historical data, and the prediction error is calculated. The historical sequence data from time t-k+1 to time t is normalized, and an unstructured text data output sequence is obtained based on the normalization. Professional information entities of the processing platform architecture are extracted from the unstructured text data, and these entities are stored in a relational database.

[0056] The other parts of this embodiment are the same as any one of the embodiments 1-5 above, so they will not be described again.

[0057] Example 7:

[0058] This embodiment further optimizes any one of the above embodiments 1-6. The pre-trained model is based on a character-level Chinese dataset. It uses a trained recognizer to perform entity recognition tasks, predicts entity strings, and obtains a list of predicted entity sets [setpre1, setpre2, ..., setpreren]. It extracts real entity strings to obtain a list of real entity sets, where setprei represents the unique set of all entities extracted from a sample.

[0059] The other parts of this embodiment are the same as any one of the embodiments 1-6 above, so they will not be described again.

[0060] Example 8:

[0061] This embodiment further optimizes any one of embodiments 1-7 above. The perception and prediction model collects multi-dimensional dynamic data throughout the entire process. Through the parallel mapping of the virtual space in the physical manufacturing system's full-domain perception, deep knowledge mining is performed to predict future data changes. Then, a trained recognizer compares the predicted data with future real data to determine if anomalies occur. Based on the data predicted by the prediction module at time t+1 and the corresponding error matrix, it determines whether the data acquired at time t+1 is abnormal. The covariance matrix is ​​calculated based on the error matrix. Different prediction and discrimination models are then learned for different network environments. Multiple indicators are used to measure the network from different perspectives. A multivariate Gaussian distribution is used to establish the relationship between these indicators, accurately detecting network anomalies and automatically adapting to different network environments to achieve proactive perception of the manufacturing system. This is combined with deep... The learning framework learns the patterns of data change, enabling the aggregation and loading of dynamic and static full-volume, multi-dimensional, and multi-scale information from the manufacturing system. This forms control strategies, which rely on automated rules to process events. It provides real-time feedback and real-time sensing data control, simulation data, and multiple data feedback mechanisms for manufacturing site reflections, major equipment failure shutdown control, missing parts alarms, broken tool alarms and decisions, and safety alarms and decisions. It constructs multi-dimensional integrated intelligent agent components across the physical, information, and business spaces of the manufacturing system. From the spatiotemporal domains, it builds intelligent agent model components for manufacturing resources, manufacturing units, and the supply chain in the intelligent manufacturing space. Around the dynamic evolution of logistics, value stream, information flow, and business flow in the manufacturing system, it establishes a dynamic collaborative operation mechanism centered on intelligent agents, realizing a real-time synchronous simulation and virtual-physical linkage control and information interaction transmission mechanism for multiple elements, businesses, and processes in the intelligent manufacturing space.

[0062] The other parts of this embodiment are the same as any one of the embodiments 1-7 above, so they will not be described again.

[0063] Example 9:

[0064] This embodiment further optimizes any one of the embodiments 1-8 above. Driven by the AI ​​brain of the manufacturing system, the algorithm optimization model synchronously collects information from the entire domain and runs intelligent agent control commands around the workshop logistics, information flow, and business flow. Through closed-loop simulation decision-making and heterogeneous integration between the manufacturing system and the AI ​​brain, the optimized control commands are transmitted to the intelligent agent model components of the manufacturing unit and the supply chain, and drive the operation of the physical manufacturing system. Through virtual-real fusion, the simulation and optimization decision-making and adaptive adjustment model of the manufacturing system are driven to achieve adaptive optimization of process parameters. The optimized parameters are fed back to the processing equipment through CNC commands to achieve closed-loop control of the entire processing process. The fault detection module uses knowledge graphs to expand the description of fault events and faulty components. It integrates and filters data existing in various fault diagnosis reports and fault case libraries, extracts useful information and organizes it as the basis for knowledge construction. The knowledge base is used to more accurately locate the system failure parts and their causes. The fault library is built by uploading files.

[0065] The other parts of this embodiment are the same as any one of the embodiments 1-8 above, so they will not be described again.

[0066] Example 10:

[0067] This embodiment is a further optimization based on any one of embodiments 1-4 above, such as... Figure 1As shown in the illustrative preferred embodiment described below, a data-driven intelligent manufacturing management method (MES) for the aerospace manufacturing industry includes: using natural language processing technology to analyze internal data, constructing a large-scale aerospace manufacturing domain knowledge base, and building a pre-trained model for the aerospace manufacturing domain on this basis. Specifically, the aerospace manufacturing domain knowledge base effectively plans and incorporates data from aircraft design and manufacturing into a unified database, incorporates data from parts manufacturing and tooling manufacturing into a tracking and management system, integrates general knowledge graphs, common sense knowledge graphs, and domain knowledge graphs, and integrates with ERP software to achieve graph retrieval and document retrieval based on the domain knowledge graph and document library. The system utilizes manufacturing information management services to process both structured and unstructured data, creating a domain knowledge graph and a domain document library. Pre-trained models, powered by an intelligent manufacturing management platform, monitor and control received and processed data variables in real-time during production, efficiently and cost-effectively performing image quality inspection. Based on the labeled model, it provides information extraction service interfaces, enabling dynamic training and deployment of labeled data. Real-time collected production parameters are transmitted to the industrial brain, where artificial intelligence algorithms perform deep learning calculations on all relevant parameters, accurately analyzing the key parameters most relevant to production quality. Parameter curve models and algorithm models are then built, using test / inspection data as the primary source, leveraging the Internet of Things and algorithm models. The technology performs real-time monitoring and detection of indicators. Using AI image processing, it extracts effective pixels and uploads a domain knowledge graph containing product defects to a cloud computing platform. Algorithm training is then conducted using deep learning and image processing techniques. Data from key production stages is aggregated and integrated via the cloud. Combining the strong fitting capabilities of convolutional neural networks and anomaly detection algorithms, deep learning is performed on massive datasets. Key features are extracted from the original images through convolution operations. Max pooling is then used after these convolution operations, treating the image as a matrix and performing matrix multiplication. After completing the convolution and pooling operations, these images are input into three fully connected layers for two full-connected operations, using different convolutional layers. The kernel trains a convolutional neural network (CNN) on a GPU, then accumulates all elements in the matrix to obtain the final result. Simultaneously, it performs deep computation and analysis on various product data to build a fault detection and perception prediction model. This model acquires demand perception of manufactured products, identifies minor faults in manufactured products, predicts measurement point faults and defects, outputs prediction results, and provides early warnings of faults and defects. Based on the industrial brain, an algorithm optimization model is built. This model is used to comprehensively analyze and evaluate process capabilities and optimize production processes, turning data into knowledge and knowledge back into data. It accurately and in real-time predicts product defect quantities and equipment failures, makes defect judgments, issues instructions to capture defective products, and provides optimal solutions for actual production by providing process parameters.

[0068] The cloud computing platform, based on the convolutional neural network algorithm for natural image recognition, divides the data to be trained into equal parts, corresponding one-to-one with the nodes on the cloud computing platform, and stores them in an evenly distributed manner. The training of the network is completed by using the data of the CNN network stored on each node. After the task receives the data, the operation module manages the large dataset of the tree diagram structure through a master node. The master node distributes the operation tasks to each sub-node. After the sub-nodes complete the data processing, they are aggregated and sent back to the master node for processing. The processed task is decomposed into multiple task modules and used on the nodes of the platform. The local changes in weights and biases are calculated through forward and backward propagation to form the intermediate key value. After all samples have been calculated, local file processing is performed. The processed local files are then aggregated with the data obtained from each training session and written to a global file.

[0069] The aerospace manufacturing knowledge base stores complete product information in the Single Product Data Source (SSPD). The SSPD itself is not a separate database; different supporting data for the product can be extracted from different usage perspectives. The aerospace manufacturing knowledge base uses the logically defined SSPD as the core access source for the underlying data and related product data of the entire aerospace manufacturing product system. The SSPD organizes product data, originally physically distributed across multiple databases, into a logically constrained single database, and establishes a knowledge system at the outset of knowledge graph creation. Each project corresponds to a unique knowledge system. The knowledge system is used as the trunk of a knowledge graph, with each file and paragraph of text within a file serving as its leaves. "Categories" and "Attributes" are dragged and dropped into the operation area. Names and types are defined for "Entity Categories" and "Attributes." Connections are made from entities to attributes, defining entity-related attributes and connecting entities from one entity to another. Relationship names are defined in the file template panel. The knowledge system is named by selecting a construction method. Data forms and sample file templates are imported, and a sample file is provided. Relationship category files and attribute category files are created according to the sample file's format. The knowledge system is located in the list, and the data files used for its creation are downloaded. "File Category," "Document Creation Time," "Knowledge System Tags," "Report Type," and whether "OCR Recognition" is required are selected. The knowledge system attribute category file and relationship category file are uploaded. The knowledge system is created through file upload and drag-and-drop. The completed knowledge system will be applied to the project, serving as the knowledge system of the project's knowledge graph. Product-related data is stored across product data distributed in different databases, forming multiple distributed databases and constructing a database that runs throughout the entire aircraft manufacturing data management system. This ensures that the product data is complete, consistent, and reliable, meeting the needs of aviation companies in all aspects, from customer selection to aircraft delivery and service support.

[0070] The file category can be selected as follows: zeroing report, flight test failure, non-conforming data, maintenance record, problem review record, and component operation record; the document creation time can be selected as the file's creation time; the knowledge system tags can be selected as the knowledge system tags associated with the file; the report type can be selected as the report type to which the file belongs; OCR optical character recognition is a system that optically scans PDF text documents, analyzes and processes the image files, and automatically recognizes and inputs the text into the computer.

[0071] When uploading a file, select the file you want to start OCR from the file management list, select "Open OCR Recognition," and enter editing mode. Once editing is complete, clicking anywhere outside the input box will end the editing process. The system constructs a multi-dimensional feature vector based on network performance metrics, normalizes each dimension, and calculates the mean and variance of each dimension in the training sequence. Then, it performs one-dimensional causal convolution and dilated convolution operations at the convolutional layer. After fusing the various channels, the uploaded data image file is obtained and automatically converted into editable text on the computer. After uploading the file, you can view the current document's recognition status in the document list. Once successfully converted, the document can be used for subsequent document information extraction and pre-annotation. During document extraction, enter the dictionary name, upload the dictionary, and select the stop word dictionary (used to remove stop words from the segmentation results during pre-annotation). Name the stop word dictionary and click the "Upload" button to upload it. Users can upload dictionaries based on sample dictionaries to enrich the extraction dictionary. Document extraction follows a structured process based on specific rules. The first step is selecting the document to extract, with only one document extracted at a time. The extraction page displays the source file on the left and the extraction results on the right, which users can edit. Clicking on an extraction result item allows editing that data. Clicking outside the input box saves the changes. Users can also click the save button below to download and save the extracted document. In project management, a knowledge graph related to knowledge and documents can be built by creating a project, uploading a knowledge system, and selecting associated documents. When creating the graph, entities, attributes, and relationships are obtained from the documents using annotation tools. The system supports importing data via triples. Clicking "Project Management" and then "Import Triples" uploads the prepared triple data, generating a complete knowledge graph based on the triples. Document annotation involves two steps: first, entity and attribute annotation, followed by relationship annotation. Completing both steps completes one annotation task. The entity annotation method is as follows: First, select a word segment in the entity annotation area, then select an entity category for the word segment. An entity card is generated for this entity in the card display area. Below the generated entity card are attribute cards, where you can select corresponding attribute names and create attribute values. Selecting an attribute name is a predefined attribute name associated with this entity within the knowledge system. Users only need to select the appropriate attribute name. In the attribute selection area, the attribute value is automatically filled in, completing the attribute value definition. This completes the entity and attribute annotation work for the project. If it is necessary to split the recognized text, select the words to be split in the word segmentation area to complete the word splitting. The system automatically associates the annotation information with the knowledge system based on the current annotation status of the project, forming a knowledge graph for the project.Users can build their own fault database by uploading files. Users enter fault information in the input box, which is then compared with the information in the fault database to determine if it is a duplicate fault. Fault information includes: product type, product name, discovery time, fault description, estimated loss, and preliminary handling status.

[0072] Graph fusion can be divided into two levels: data pattern fusion and data fusion. Data pattern fusion mainly involves pattern mapping and standardization. Together, they form the information hub system for controlling the efficient operation of the manufacturing system, serving as the core of intelligent decision-making and control in the intelligent manufacturing system. The information hub system receives information from the entire manufacturing system's perception domain, processes it, and outputs control commands, or stores it as knowledge. The entire manufacturing system's state information, through digital virtual-physical mapping, is transmitted to the AI ​​brain. A large number of AI algorithm models for different scenarios are aggregated within the manufacturing system's information hub and transmitted to the physical manufacturing system, forming the manufacturing system's analysis and decision-making center. This enables the virtual transmission, storage, and processing of information to generate various system decision information, governing and controlling the operation of the manufacturing system, ensuring its overall coordinated operation, and achieving optimized operation.

[0073] Domain knowledge graphs obtain entity names through the NER model. The NER model first determines the required entity names (mentions), defining digital information about the geometry, topology, materials, processes, and relevant technical specifications of the manufactured products, storing this information in the Engineering Data Set and the Automatic Parts List (APL), respectively. Then, each mention is matched to its corresponding candidate entities in the knowledge graph, sorted, aggregated, and integrated with multi-dimensional, heterogeneous manufacturing process data. Sequence labeling is performed on a character-by-character basis for modeling. The pre-trained model receives training data in the form of text data and dictionary tuples, updating the existing model based on context to train NER. The model iterates through the training data, adding a smart progress meter (tqdm) before the dataloader to obtain sufficient iterations. Tqdm is then applied to deep learning, using the iterative progress bar (tqdm()) function to create a progress bar and save training process information. The pre-trained model processes the maximum activation feature maps of the convolutional layers during training, then occludes the input image. The occluded new image is then used as new input to the network for further training. To accelerate training in image recognition, images are pre-processed to remove redundant and interfering information. The processed image is then used to segment recognition regions. Simultaneously, the latest historical data is used to predict the state at the next time step, and the prediction error is calculated. The historical sequence data from time t-k+1 to time t is normalized, resulting in an unstructured text output sequence. Professional information entities representing the processing platform architecture are extracted from this unstructured text data and stored in a relational database. The dictionary includes start and end indices for named entities in both the text and category. The pre-trained language model can preprocess images and run on all documents, extracting entities and storing them separately. The next time a user searches for a word, the search term will be matched against a smaller list of entities in each document. Named Entity Recognition (NER) is a natural language processing technique used to extract appropriate entities from given text content and classify the extracted entities into predefined categories. The pre-trained model is based on a character-level Chinese dataset. It uses a trained recognizer to perform entity recognition tasks, predicting entity strings to obtain a list of predicted entity sets [setpre1, setpre2, ..., setpreren]. Then, it extracts real entity strings to obtain a list of real entity sets, where setprei represents the unique set of all entities extracted from a sample. There are three main entity types: location, organization, and person. The data is labeled using the BIO style.

[0074] The perception and prediction model collects multi-dimensional dynamic data throughout the entire process. Through parallel mapping of the virtual space within the physical manufacturing system, it performs deep knowledge mining to predict future data changes. Then, a trained recognizer compares the predicted data with future real data to determine if anomalies exist. Based on the data predicted by the prediction module at time t+1 and the corresponding error matrix, it judges whether the data acquired at time t+1 is abnormal. The covariance matrix is ​​calculated from the error matrix, and different prediction and discrimination models are learned for different network environments. Multiple indicators are used to measure the network from different perspectives, and a multivariate Gaussian distribution is used to establish the relationship between these indicators, accurately detecting network anomalies and automatically adapting to different network environments to achieve proactive perception of the manufacturing system. Combined with a deep learning framework, it learns the data change patterns. The system aims to aggregate and load dynamic and static information from the manufacturing system across all dimensions and scales, forming control strategies. It relies on automated rules to process events, providing real-time feedback and real-time sensing data control, simulation data, and multiple data feedback mechanisms for on-site reflections, equipment downtime control due to major malfunctions, missing parts alarms, broken tool alarms and decisions, and safety alarms and decisions. This constructs a multi-dimensional integrated intelligent agent component encompassing the physical, information, and business spaces of the manufacturing system. From the spatiotemporal domains, it builds intelligent agent model components for manufacturing resources, manufacturing units, and the supply chain within the intelligent manufacturing space. Around the dynamic evolution of logistics, value streams, information flows, and business flows in the manufacturing system, it establishes a dynamic collaborative operation mechanism centered on intelligent agents, enabling real-time synchronous simulation and virtual-physical linkage control and information exchange transmission mechanisms for multiple elements, businesses, and processes within the intelligent manufacturing space.

[0075] In actual machining, the perception and prediction model perceives the real-time status of the physical equipment's machining process through input and target sequences. Facing the control commands of the virtual and physical manufacturing system, it constructs a global perception model that preprocesses images. The model transmits real-time data on equipment status (machine tool vibration, power, axis speeds, feed rates, current running program, etc.), tool status (tool life, cutting force, model, etc.), operating conditions (coolant, temperature, etc.), and process quality (on-machine inspection process quality data, etc.) to the global perception model. Simulation data and historical machining data are then fed back to the machine tool intelligent agent model, achieving closed-loop control of the manufacturing process.

[0076] The algorithm optimization model, relying on the high computing power requirements of artificial intelligence algorithms, constructs a multi-level manufacturing system intelligent agent model component capability model and a large-scale cross-organizational resource collaborative optimization scheduling model through a virtual space data aggregation and integration channel. The optimization scheduling model calls upon the artificial intelligence algorithm model established in the AI ​​brain, aggregates real-time operating status data of the manufacturing system, performs computational analysis and solutions, and inputs the results into the algorithm model for simulation analysis. Utilizing remote control commands, parameter corrections, and message transmission from the actuators, it reacts to the physical manufacturing system, aggregating, analyzing, fusing, and iteratively optimizing the entire domain data of the physical manufacturing system, and establishing a manufacturing system resource and task... The process collaboration optimization mechanism optimizes and controls the entire aerospace manufacturing system based on the dynamic optimization and intelligent decision-making effect. It receives production tasks, plans, schedules, and dynamically schedules equipment, materials, and tools, and allocates NC program manufacturing resources in real time to complete the dynamic allocation of various manufacturing elements. The control commands are converted through the neural center to form driving control of each element agent, complete the simulation operation, and feed the simulation results back to the AI ​​control center to achieve closed-loop optimization. Each element agent generates driving control commands to drive the operation of the physical manufacturing system, and establishes a fusion model of multi-dimensional and multi-scale manufacturing process data dynamic evolution information expression and fusion driving and system state information. The fusion model establishes an interaction and fusion mechanism between physical and virtual manufacturing systems, integrating technologies such as the Internet of Things, deep learning, and image recognition. This enables the fusion of multi-source, multi-dimensional, and heterogeneous quality data from the manufacturing site and virtual simulation. Real-time monitoring data is dynamically updated, and algorithmic optimization models predict the equipment manufacturing process status. This predictive and decision-making control of the entire process flow facilitates communication and information exchange between agents. Through a message service platform, agents can modify their own status information to update it or obtain status information from other subscribed agents to achieve information exchange. This allows for dynamic control of the manufacturing process. Through dynamic optimization control and operational decisions across the entire process, data fusion and data-driven approaches are achieved across all elements of the manufacturing process. Resource collaborative optimization is also implemented in complex and uncertain environments, thereby improving the technological capabilities of product manufacturing. This reduces the difficulty and cost of factory production management and increases equipment utilization and production efficiency.

[0077] Driven by the AI ​​brain of the manufacturing system, the algorithm optimization model synchronously collects information from the entire domain and runs intelligent agent control commands around the workshop logistics, information flow, and business flow. Through closed-loop simulation decision-making and heterogeneous integration between the manufacturing system and the AI ​​brain, the optimized control commands are transmitted to the intelligent agent model components of the manufacturing unit and supply chain, and drive the operation of the physical manufacturing system. Through virtual-real fusion, the simulation and optimization decision-making and adaptive adjustment model of the manufacturing system are driven to achieve adaptive optimization of process parameters. The optimized parameters are fed back to the processing equipment through CNC commands, realizing closed-loop control of the entire processing process, improving processing accuracy and efficiency, as well as the continuous improvement of the explicit and implicit capabilities of the manufacturing system and the spiral evolution of the production efficiency of the parts manufacturing system.

[0078] The fault detection module utilizes knowledge graphs to expand the descriptions of fault events and faulty components. It integrates and filters data from various fault diagnosis reports and fault case libraries, extracting useful information and organizing it as the foundation for knowledge construction. This knowledge base is used to more accurately locate system failure points and their causes, and a fault database is built by uploading files. The fault detection module compares the information with the fault database to determine if it is a repeat fault. Fault information includes: product type, product name, discovery time, fault description, estimated loss, and preliminary handling status.

[0079] like Figure 2As shown, the perception prediction model constructs manufacturing product projects. Each project contains a knowledge graph. Based on the current annotation status of the project, the annotation information is automatically associated with the knowledge system to form the knowledge graph of the project. The number of documents associated with the current project is displayed. Entity and attribute categories and relationship categories are added to the graph of the current project. The knowledge system is constructed by importing data forms. After the knowledge system is created, the knowledge graph is created. Each project corresponds to a unique knowledge system. The created knowledge system will be applied to the project as the knowledge system of the project's knowledge graph. After the knowledge system is created, the created knowledge system is found in the knowledge system list. The knowledge system attribute category file and knowledge system relationship category file are uploaded respectively. The dictionary / model knowledge system is uploaded to the document management library and the knowledge graph management library. Through graph editing and graph query, the knowledge system visualization and knowledge graph visualization are completed. By selecting documents, pre-annotating, assigning tasks, and annotating entities, relationship and attribute annotations are performed, and triple data is imported into the knowledge graph to form the processing platform architecture. The text annotation platform architecture is divided into a data layer, a model layer, a data management layer, a model algorithm layer, an expression layer, a view layer, an application layer, and a user layer. The data layer primarily handles text data acquisition for annotation tasks, capturing tag list information, integrating this information into tag prediction, enabling dynamic training and deployment of annotation data, and providing information extraction services based on the annotation model. Data of various structures within the data layer undergoes information extraction techniques such as text extraction, entity recognition, and relation extraction from the model layer. Combined with a knowledge representation system, this forms highly structured triplet data. This data can be stored in a graph database and displayed in a clear and intuitive manner. Combined with data statistics techniques and graph algorithms, the stored graph data is visualized, supporting queries for shortest paths, nodes within n hops, and other information to deeply mine implicit relationships between entities. Different types of data can be exported in different file formats.The visualization interface can be saved as an image, and the corresponding triplet data can be exported in CSV and JSON formats. The data management layer mainly implements the parsing, processing, storage, updating and maintenance of user-uploaded data, integrating the information of the entire output word vector of a sequence into the annotation model. The view layer provides users with annotation functions such as word segmentation annotation, entity annotation, and relation annotation. Building on the view layer, the expression layer provides deeper application services such as intelligent question answering and intelligent retrieval, deeply mining the features and relationships between data to promote industry intelligence. According to the annotation task, it obtains document data content such as user dictionaries, entity category definitions, and relation category definitions, directly connecting to the fully connected layer, integrating character information into the word-level temporal recurrent neural network LSTM, and using the cross-entropy loss function for learning. Based on the annotation model, it provides information extraction related service interfaces, inputting text sequences. The sequence encoding is performed in a bidirectional temporal recurrent neural network (LSTM). The output encoding is then transformed using a fully connected layer, and the prediction task is performed with the number of labels as the dimension of the target neurons. The model algorithm layer mainly builds a labeling model based on information extraction techniques combined with user-annotated data. It uses algorithms such as relation extraction to achieve word segmentation, dependency parsing, and entity recognition. It can also be extended to a rich application layer. The output is used as the emission score matrix of the Conditional Random Field (CRF) layer and combined with the CRF layer. A function log-sum-exp is defined, using the log_sum_exp score of the CRF as the objective function. The transition score matrix is ​​initialized using some initialization methods, and the parameters are updated through gradient descent of the neural network. The application layer abstracts various basic services and encapsulates them into interfaces to facilitate the acceptance of updated information from the graph database and provides a user-friendly visualization interface to the labeling users.

[0080] The other parts of this embodiment are the same as any one of the embodiments 1-4 above, so they will not be described again.

[0081] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications or equivalent changes made to the above embodiments based on the technical essence of the present invention shall fall within the protection scope of the present invention.

Claims

1. A data-driven intelligent manufacturing management method for the aerospace manufacturing industry, characterized in that: include: This paper utilizes natural language processing (NLP) technology to analyze internal data and construct a large-scale knowledge base for the aerospace manufacturing field. Based on this, a pre-trained model for the aerospace manufacturing field is built. The key features include: effective planning of aircraft design and manufacturing data within a unified database; integration of parts manufacturing and tooling manufacturing data into a tracking and management system; fusion of general knowledge graphs, common-sense knowledge graphs, and domain knowledge graphs; and integration with ERP software. This enables manufacturing information management based on domain knowledge graphs and document libraries, providing graph retrieval and document retrieval application services. Simultaneously, it processes structured and unstructured data to form domain knowledge... Image recognition and domain documentation libraries are provided. Pre-trained models utilize the Industrial Brain platform to monitor and control received and processed data variables in real time during production, efficiently and cost-effectively completing image quality inspection. Based on the labeled model, information extraction service interfaces are provided, enabling dynamic training and deployment of labeled data. Real-time collected production parameters are transmitted to the Industrial Brain, where artificial intelligence algorithms perform deep learning calculations on all related parameters, accurately analyzing the key parameters most relevant to production quality. Parameter curve models and algorithm models are built, using test / inspection data as the main body, leveraging IoT and algorithm model technology for real-time monitoring and detection of indicators. AI image technology is used to improve... Valid pixels are extracted, and the domain knowledge graph containing product defects is uploaded to a cloud computing platform. Algorithm training is performed using deep learning and image processing technologies. Data from key production stages is aggregated and integrated via the cloud. Combining the strong fitting ability and anomaly detection algorithms of convolutional neural networks, deep learning is applied to massive datasets. Key features are extracted from the original images through convolution operations. Max pooling is then used after the convolution operations, treating the image as a matrix and performing matrix multiplication. After completing the convolution and pooling operations, these images are input into three fully connected layers for two full-connection operations. The convolutional neural network is trained on a GPU using different convolutional kernels. The system uses a CNN to accumulate all elements in the matrix to obtain the final result. Simultaneously, it performs deep computation and analysis on various product data to build a fault detection and perception prediction model. This model acquires demand perception of manufactured products, identifies minor faults, predicts measurement point faults and defects, outputs prediction results, and provides early warnings of faults and defects. Based on an industrial brain, an algorithm optimization model is built. This model is used for comprehensive analysis and evaluation of process capabilities and optimization of production processes. Data is transformed into knowledge, and knowledge is then transformed back into data. This allows for accurate and real-time prediction of product defect quantities and equipment failures, defect determination, issuing instructions to capture defective products, and providing optimal solutions for actual production based on process parameters.

2. The intelligent manufacturing management method for the aerospace manufacturing industry based on the knowledge-based brain data, as described in claim 1, is characterized in that... include: The cloud computing platform, based on the convolutional neural network algorithm for natural image recognition, divides the data to be trained into equal parts, corresponding one-to-one with the nodes on the cloud computing platform, and stores them in an evenly distributed manner. The training of the network is completed by using the data of the CNN network stored on each node. After the task receives the data, the operation module manages the large dataset of the tree diagram structure through a master node. The master node distributes the operation tasks to each sub-node. After the sub-nodes complete the data processing, they are aggregated and sent back to the master node for processing. The processed task is decomposed into multiple task modules and used on the nodes of the platform. The local changes in weights and biases are calculated through forward and backward propagation to form the intermediate key value. After all samples have been calculated, local file processing is performed. The processed local files are then aggregated with the data obtained from each training session and written to a global file.

3. The intelligent manufacturing management method for the aerospace manufacturing industry based on the knowledge-based brain data, as described in claim 1, is characterized in that... include: Product data that was originally physically distributed across multiple databases is organized into a single, logically constrained database, resulting in a logically single product data source (SSPD). The logically single product data source SSPD is used as the core underlying data source for all related product data in the entire aerospace manufacturing product system. A knowledge system is created first when building the knowledge graph, with each project corresponding to a unique knowledge system. This knowledge system serves as the trunk of the knowledge graph, and each file and its text paragraphs are considered its leaves. "Categories" and "Attributes" are dragged into the operation area, and names and types are defined for "Entity Categories" and "Attributes." Connections are established from entities to attributes, defining relationship names to associate entities with attributes. Connections are also established from one entity to another, and relationship names are defined in the file template panel. The knowledge system is named by selecting the method for building the knowledge system, importing data forms and sample file templates, providing a sample file, and creating relationships according to the sample file's format. Create attribute category files according to the format in the example file. In the knowledge system list, find the knowledge system you are creating, download the data files used to create the knowledge system, select "File Category", "Document Creation Time", "Knowledge System Tags", "Report Type" and whether "OCR Recognition" is required, and upload the knowledge system attribute category files and knowledge system relationship category files respectively. Create the knowledge system by uploading files and dragging and dropping. The completed knowledge system will be applied to the project as the knowledge system of the project's knowledge graph. It will store product-related data between product data distributed in different databases, forming multiple distributed databases, and building a system database that runs through all data sources in the entire process of aircraft manufacturing data management to meet the needs of aviation enterprises from customer selection to aircraft delivery and service support.

4. The intelligent manufacturing management method for knowledge-based brain data in the aviation manufacturing industry according to claim 1, characterized in that, include: The system constructs a multi-dimensional feature vector based on network condition metrics, normalizes each dimension, and calculates the mean and variance of each dimension in the training sequence. Then, it performs convolution operations using one-dimensional causal convolution and dilated convolution at the convolutional layer. After fusing the various channels, the uploaded data image file is obtained and automatically converted into editable text on the computer. When uploading a file, select the file to be started in the file management list, select to open OCR recognition, enter the editing state, and click any area outside the input box to end the editing state. After uploading the file, view the recognition status of the current document in the document list.

5. The intelligent manufacturing management method for the aerospace manufacturing industry based on the knowledge-based brain data, as described in claim 1, is characterized in that... This includes: obtaining a complete map based on triplet data through annotation tools when creating the map, and adding data by importing triplet data; Specifically, this involves: clicking Project Management - Import Triples, uploading the prepared triplet data, and then generating a complete knowledge graph based on the triplet data; document annotation is divided into two steps: first, complete the entity and attribute annotation, and then annotate the relationships. Once both steps are completed, an annotation task is finished; the entity annotation method is as follows: first, select a word segmentation in the entity annotation area, then select the entity category after word segmentation, and then generate the corresponding entity card for the entity in the card display area. Below the generated entity card is the attribute card, where you select the corresponding attribute name and create attribute values; select the attribute name, which is the attribute name defined in the knowledge system and associated with the entity; the system automatically associates the annotation information with the knowledge system based on the current annotation status of the project, forming the domain knowledge graph of the project.

6. The intelligent manufacturing management method for the aerospace manufacturing industry based on the knowledge-based brain data, as described in claim 5, is characterized in that... include: The domain knowledge graph is processed by a Named Entity Recognition (NER) model based on a pre-trained language model to obtain entity names; The NER model first identifies the required entity names (mentions), defines the digital information of the manufactured product's geometry, topology, materials, processes, and relevant technical specifications, and stores these in the Engineering Data Set and Automatic Parts List (APL), respectively. Then, each mention is matched to its corresponding candidate entities in the knowledge graph, sorted, aggregated, and integrated multi-dimensional and heterogeneous manufacturing process data, and modeled using character-based sequence labeling. The pre-trained model receives training data in the form of text data and dictionary tuples, updates the existing model based on context to train NER, and iterates on the training data. A smart progress meter (tqdm) is added before the data loader of the iterator to obtain sufficient iterations. The tqdm is then applied to deep learning, and an iterative progress bar (tqdm()) function is used to create a progress bar to save training process information.

7. The intelligent manufacturing management method for the aerospace manufacturing industry based on the knowledge-based brain data, as described in claim 1, is characterized in that... include: During training, the pre-trained model processes the maximum activation feature map of the convolutional layer and then occludes the input image. The occluded new image is then used as the new input to the network to continue training the model. In the image recognition process, in order to further accelerate the training speed, the image is preprocessed to remove redundant and interference information. Then, the processed image is divided into recognition regions. At the same time, the latest historical data is used to predict the data and prediction error of the next time step. After normalizing the historical sequence data from time t-k+1 to time t, an unstructured text data output sequence is obtained based on the normalization. Professional information entities of the processing platform architecture are extracted from the unstructured text data and these entities are stored in a relational database.

8. The intelligent manufacturing management method for the aerospace manufacturing industry based on the knowledge-based brain data according to claim 1, characterized in that, include: The pre-trained model is based on a character-level Chinese dataset. It uses a trained recognizer to perform entity recognition tasks, predicts entity strings, and obtains a list of predicted entity sets [setpre1, setpre2, ..., setpreren]. It then extracts real entity strings to obtain a list of real entity sets, where setprei represents the unique set of all entities extracted from a sample.

9. The intelligent manufacturing management method for knowledge-based brain data in the aviation manufacturing industry according to claim 1, characterized in that, include: The perception and prediction model collects multi-dimensional dynamic data throughout the entire process. Through parallel mapping of the virtual space within the physical manufacturing system, it performs deep knowledge mining to predict future data changes. Then, a trained recognizer compares the predicted data with future real data to determine if anomalies exist. Based on the data predicted by the prediction module at time t+1 and the corresponding error matrix, it determines whether the data acquired at time t+1 is abnormal. The covariance matrix is ​​calculated from the error matrix, and different prediction and discrimination models are learned for different network environments. Multiple indicators are used to measure the network from different perspectives, and a multivariate Gaussian distribution is used to establish the relationship between these indicators, accurately detecting network anomalies and automatically adapting to different network environments to achieve proactive perception of the manufacturing system. Combined with a deep learning framework, it learns the data change patterns. The system aims to aggregate and load dynamic and static information from the manufacturing system across all dimensions and scales, forming control strategies. It relies on automated rule-based events for processing, providing real-time feedback and real-time sensing data control, simulation data, and multiple data feedback mechanisms for manufacturing site reflections, major equipment failure shutdown control, missing parts alarms, broken tool alarms and decisions, and safety alarms and decisions. It constructs multi-dimensional integrated intelligent agent components encompassing the physical, information, and business spaces of the manufacturing system. From the spatiotemporal domains, it builds intelligent agent model components for manufacturing resources, manufacturing units, and the supply chain within the intelligent manufacturing space. Around the dynamic evolution of logistics, value stream, information flow, and business flow in the manufacturing system, it establishes a dynamic collaborative operation mechanism centered on intelligent agents, realizing a real-time synchronous simulation and virtual-physical linkage control and information interaction transmission mechanism for multiple elements, businesses, and processes within the intelligent manufacturing space.

10. The intelligent manufacturing management method for knowledge-based brain data in the aviation manufacturing industry according to claim 1, characterized in that, include: Driven by the AI ​​brain of the manufacturing system, the algorithm optimization model synchronously collects information across the entire domain and executes intelligent agent control commands around the workshop logistics, information flow, and business flow. Through closed-loop simulation decision-making and heterogeneous integration between the manufacturing system and the AI ​​brain, the optimized control commands are transmitted to the intelligent agent model components of the manufacturing unit and supply chain, driving the operation of the physical manufacturing system. Through virtual-real fusion, the simulation and optimization decision-making of the manufacturing system and the adaptive adjustment model achieve adaptive optimization of process parameters. The optimized parameters are fed back to the processing equipment through CNC commands, realizing closed-loop control of the entire processing process. The fault detection module uses knowledge graphs to expand the description of fault events and faulty components. It integrates and filters data from various fault diagnosis reports and fault case libraries, extracts useful information, and organizes it as the basis for knowledge construction. The knowledge base is used to more accurately locate the system failure parts and their causes. By uploading files to build a fault library, the fault detection module compares the information with the fault library to determine whether it is a repeat fault. The fault information includes: product type, product name, discovery time, fault description, estimated loss, and preliminary handling status.

Images