Intelligent clothing industry production control method and system based on data analysis

Abstract

The present invention provides a production control method and system for an intelligent clothing industry based on data analysis, which relates to the field of intelligent manufacturing technology, and includes: setting an intelligent data acquisition platform, collecting heterogeneous data for common feature extraction, generating a comprehensive data set, and constructing a digital twin model for clothing production; performing data extraction based on the comprehensive data set, generating triples and constructing a clothing production knowledge graph, constructing a clothing production data analysis model, extracting key features and determining implicit association rules, generating a cross-modal knowledge graph, identifying clothing production control rules, and solving clothing production control parameters; constructing a decision evaluation model to parse clothing production control parameters into production control decisions, perform evaluation, dynamically correct based on the evaluation results, obtain optimized production control decisions, send them to equipment terminals, adaptively optimize clothing production, collect real-time production data, update the clothing production digital twin model and the clothing production knowledge graph, and periodically optimize until the production task is completed.

Description

Technical Field

[0001] The present invention relates to the field of intelligent manufacturing technology, and in particular to a production control method and system for the intelligent clothing industry based on data analysis. Background Art

[0002] The clothing industry is a traditional labor-intensive industry. The production process involves design, sampling, cutting, sewing, ironing, packaging and other links. Traditional clothing production control mainly relies on manual experience and simple statistical methods, and there are problems such as low production efficiency, unstable quality, and waste of resources. With the development of information technology and the intensification of market competition, clothing companies are in urgent need of introducing intelligent production control methods to improve production efficiency, reduce costs, and meet diversified and personalized market needs.

[0003] At present, domestic and foreign scholars and enterprises have begun to explore intelligent clothing production control methods. Most of them use machine learning algorithms, such as support vector machines and decision trees, to model and predict production data. However, existing research and applications lack system optimization of the entire clothing production process, the collection and processing of production data is not intelligent enough, and the knowledge-driven decision support is not in-depth enough.

[0004] Therefore, a method is urgently needed to solve the problems existing in the prior art. Summary of the invention

[0005] The embodiments of the present invention provide a method and system for production control in the intelligent clothing industry based on data analysis, which can at least solve some of the problems existing in the prior art.

[0006] A first aspect of an embodiment of the present invention provides a production control method for the smart clothing industry based on data analysis, comprising:

[0007] An intelligent data collection platform is set up, and based on the data collection agent module in the intelligent data collection platform, heterogeneous data in clothing production equipment, supply chain management system and sales terminal system are collected in combination with data stream real-time transmission technology. For the heterogeneous data, missing values ​​and outliers are identified through rule-based data cleaning algorithms and learning-based data cleaning algorithms, and the data are cleaned and converted into vectorized representations. Common features are extracted through a multi-view data fusion algorithm based on tensor decomposition, and a comprehensive data set is generated in combination with the vectorized representation. Multimodal feature representations are extracted through a multimodal feature representation learning algorithm and a digital twin model of clothing production is constructed;

[0008] The core concepts, attributes and relationships in the field of clothing production are defined through the ontology knowledge modeling method, data extraction is performed based on the comprehensive data set, triples are generated and a clothing production knowledge graph is constructed, a clothing production data analysis model is constructed based on the graph attention network algorithm, the clothing production knowledge graph is added to the clothing production data analysis model, key features are adaptively extracted and implicit association rules are determined through the attention mechanism, a cross-modal knowledge graph is generated, the clothing production control rules are identified through the knowledge reasoning algorithm in combination with the prior knowledge corresponding to the clothing production knowledge graph, and the clothing production control parameters are solved in combination with the multi-objective optimization algorithm;

[0009] A decision evaluation model is constructed based on a reinforcement learning algorithm and a reward function is set. The clothing production control parameters are parsed into production control decisions. The production control decisions are evaluated through the decision evaluation model. Dynamic corrections are made based on the evaluation results to obtain optimized production control decisions. The optimized production control decisions are sent to equipment terminals in the clothing production workshop and clothing production is adaptively optimized in combination with task requirements. Real-time production data is collected during the production process and the clothing production digital twin model and the clothing production knowledge graph are updated. Periodic optimization is performed until the production task is completed.

[0010] In an optional embodiment,

[0011] An intelligent data collection platform is set up, and based on the data collection agent module in the intelligent data collection platform, heterogeneous data in clothing production equipment, supply chain management system and sales terminal system are collected in combination with data stream real-time transmission technology. For the heterogeneous data, missing values ​​and abnormal values ​​are identified through rule-based data cleaning algorithms and learning-based data cleaning algorithms, and the data are cleaned and converted into vectorized representations. Common features are extracted through a multi-view data fusion algorithm based on tensor decomposition, and a comprehensive data set is generated in combination with the vectorized representation. The multi-modal feature representation is extracted through a multi-modal feature representation learning algorithm and a digital twin model of clothing production is constructed, including:

[0012] An intelligent data collection platform is set up based on a microservice architecture, wherein the intelligent data collection platform includes a data collection agent module, a message middleware, a data cleaning module and a digital twin engine. The data collection agent module adopts a sidecar mode and is deployed in a containerized manner with the data interfaces of clothing production equipment, supply chain management system and sales terminal system. For the data in the clothing production equipment, equipment parameters and status data are periodically collected. For non-standard private protocols, the edge computing gateway is combined to generate standard equipment data through protocol parsing and data conversion and send it to the message middleware. For the supply chain management system, it is connected through a descriptive status interface, and purchase orders, inventory levels and logistics status are regularly pulled as supply chain data. The supply chain data is extracted into a pre-set data mart and sent to the message middleware. For the sales terminal system, the sales quantity of the e-commerce platform is regularly obtained through data crawling technology. For unstructured user comment data, product attribute words and sentiment polarity are extracted in combination with sentiment analysis and topic mining technology and sent to the message middleware. The data in the message middleware is combined to obtain the heterogeneous data.

[0013] For the structured data in the heterogeneous data, the integrity and conflict of the data are verified by a data quality verification function based on a distributed structured query language engine, and non-compliant data is identified and removed in combination with business rules. For unstructured data, a data preprocessing algorithm based on a distributed machine learning library is used to map words in text data to a low-dimensional vector space using a word embedding algorithm, generate word vectors and perform cleaning, and multi-scale textures and edges in image data are extracted through a convolutional neural network to generate image feature vectors, which are cleaned according to a pre-set quality threshold. For time series data, a symbolic aggregation approximation algorithm is selected for segmented aggregation forcing and a symbolic mapping algorithm is used to map the time series data into a symbolic sequence, reduce the data dimension and extract the morphological features of the time series, and perform outlier identification and missing value completion based on the morphological features;

[0014] During the cleaning process, the thresholds of missing values ​​and outliers are adjusted according to data quality through adaptive threshold strategies. For data that cannot be judged during the cleaning process, the most representative data is selected for annotation through artificial intelligence technology. The cleaning is repeated and the cleaned heterogeneous data is converted into vectorized representation.

[0015] For heterogeneous data represented by vectorization, a high-order tensor is constructed through a multi-perspective data fusion algorithm based on tensor decomposition, wherein each dimension of the high-order tensor corresponds to a data perspective, and the high-order tensor is low-rank approximated through the Tucker decomposition algorithm to obtain a core tensor and multiple factor matrices, and domain ontology knowledge is applied as constraints to the core tensor and the factor matrix. A heterogeneous information network is constructed through a graph-based data fusion strategy, and feature vectors of different perspectives are used as nodes, and the correlation between features is used as edges. The weight of the edge is determined according to the semantic similarity, and the low-dimensional embedding representation of the network nodes is learned by combining the random walk network embedding algorithm and mapped to the same semantic space to obtain common features and construct the comprehensive data set based on the common features, and the multimodal feature representation in the comprehensive data set is extracted through the multimodal feature representation learning algorithm, and a virtual simulation environment of a clothing production workshop is constructed in combination with the digital twin engine, and the entity is abstracted into a three-dimensional model in combination with a pre-set three-dimensional collection library to obtain the clothing production digital twin model.

[0016] In an optional embodiment,

[0017] The high-order tensor is approximated to a low rank using the Tucker decomposition algorithm as shown in the following formula:

[0018] ;

[0019] in, X represents a high-order tensor, G represents the core tensor, × n represents the modular product operation of tensors, Indicates N factor matrix, N represents the total number of factor matrices, n represents the index of the factor matrix, λ represents the regularization parameter for the smoothness constraint, tr represents the trace of the matrix, L n represents the graph Laplacian matrix, represents the factor matrix, ‖‖ F represents the Frobenius norm.

[0020] In an optional embodiment,

[0021] The core concepts, attributes and relationships in the field of clothing production are defined through the ontology knowledge modeling method, data extraction is performed based on the comprehensive data set, triples are generated and a clothing production knowledge graph is constructed, a clothing production data analysis model is constructed based on the graph attention network algorithm, the clothing production knowledge graph is added to the clothing production data analysis model, key features are adaptively extracted and implicit association rules are determined through the attention mechanism, a cross-modal knowledge graph is generated, and the clothing production control rules are identified through the knowledge reasoning algorithm in combination with the prior knowledge corresponding to the clothing production knowledge graph, and the clothing production control parameters are solved in combination with the multi-objective optimization algorithm, including:

[0022] Based on expert interviews and literature research, the core concepts, attributes and relationships in the field of clothing production are obtained, and the knowledge in the field of clothing production is formally represented. Based on the ontology knowledge modeling method, the hierarchical structure is determined, and the superordinate concepts and subordinate concepts are divided. The acquired knowledge is integrated and the production task ontology is defined. The production process is abstracted into a directed acyclic graph, in which nodes represent production tasks and edges represent dependencies between tasks. The semantic mapping relationship between different ontologies is identified through an ontology matching algorithm based on semantic description, and knowledge is fused based on the semantic mapping relationship. The ontology alignment algorithm based on logical reasoning is combined to eliminate semantic objections and perform ontology reasoning and expansion to obtain an ontology knowledge base.

[0023] The comprehensive data set is semantically annotated and knowledge extracted, entities, attributes and relationships are extracted to generate triples, the triples are imported into a preset graph database, attribute indexes of vertices and edges are constructed, the clothing production knowledge graph is obtained, a clothing production data analysis model is constructed based on a graph attention network algorithm, the clothing production knowledge graph is converted into a graph embedding representation and added to the clothing production data analysis model, the semantic representation of vertices in the clothing production knowledge graph is learned through a graph attention network, neighbor information is aggregated through attention weights and its own feature representation is updated, the importance weights of elements in the clothing production knowledge graph are calculated based on preset attention functions of different granularities and the key features are weighted to generate the key features, the key features are interpreted, the frequent co-occurrence patterns and sequence patterns between production factors are determined in combination with an association rule mining algorithm, the semantic association structure is identified in combination with a graph pattern matching technology, the implicit association law is obtained, and the cross-modal knowledge graph is constructed based on the implicit association law and the clothing production knowledge graph;

[0024] Prior knowledge corresponding to the clothing production knowledge graph is obtained, the prior knowledge is integrated with the cross-modal knowledge graph, and the cross-modal knowledge graph integrated with the prior knowledge is logically inferred according to the existing production rules through a knowledge reasoning algorithm to obtain the clothing production control rules. Based on a multi-objective optimization algorithm, an objective function is constructed with the goals of maximizing production efficiency and minimizing quality loss. The objective function is solved based on a genetic algorithm to generate a Pareto optimal solution set and perform optimization search through taboo search to obtain the clothing production control parameters.

[0025] In an optional embodiment,

[0026] Based on the multi-objective optimization algorithm, the objective function is constructed with the goal of maximizing production efficiency and minimizing quality loss as shown in the following formula:

[0027] ;

[0028] in, x i Indicates production tasks i The processing time, t i Indicates the task i Standard processing time, c i Indicates the task i The mass loss coefficient, μ i Indicates production tasks i The mass loss rate, m Indicates the total number of tasks, represents the actual value of mass loss, Indicates the actual value of production efficiency.

[0029] In an optional embodiment,

[0030] A decision evaluation model is constructed based on a reinforcement learning algorithm and a reward function is set. The clothing production control parameters are parsed into production control decisions. The production control decisions are evaluated through the decision evaluation model. Dynamic corrections are made based on the evaluation results to obtain optimized production control decisions. The optimized production control decisions are sent to equipment terminals in the clothing production workshop and clothing production is adaptively optimized in combination with task requirements. Real-time production data in the production process is collected and the clothing production digital twin model and the clothing production knowledge graph are updated. Periodic optimization is performed until the production task is completed, including:

[0031] Based on the reinforcement learning algorithm, according to the characteristics of clothing production regulation and control decisions, the clothing production regulation parameters are used as state variables to construct a multidimensional state space, the clothing production regulation decisions are used as action variables to construct an action space, a multi-objective reward function is set according to production efficiency, quality loss and energy consumption, and the decision evaluation model is constructed based on the multi-dimensional state space, the action space and the multi-objective reward function;

[0032] The clothing production control parameters are parsed into production control decisions and added as input data to the decision evaluation model, the expected cumulative reward corresponding to the current production control decision is calculated through forward propagation, and the expected cumulative reward is compared with a preset reward threshold. If the expected cumulative reward is greater than the reward threshold, the current production control decision is considered to be an optimized production control decision. Otherwise, the production control decision is optimized through a policy gradient algorithm to obtain an optimized production control decision.

[0033] The optimized production control decision is sent to the equipment terminal of the clothing production workshop. The equipment terminal adaptively optimizes the clothing production process according to the received decision instructions and the production task requirements, monitors the clothing production status in real time and dynamically adjusts the process parameters and resource allocation, collects real-time production data under different production states, transmits the real-time production data to the clothing production digital twin model, updates it through data synchronization and mapping, generates simulation data through the updated clothing production digital twin model, and updates the clothing production knowledge graph and the cross-modal knowledge graph;

[0034] An optimization cycle is set and the optimization is performed cyclically. In each optimization cycle, the optimized production control decision is repeatedly executed and generated until the production task is completed or the preset optimization goal is reached.

[0035] In an optional embodiment,

[0036] The production control decision is optimized by the policy gradient algorithm as shown in the following formula:

[0037] ;

[0038] in, Indicates in status s Next, the current strategy π θ Select Action a The probability of Indicates in status s Next, the old strategy π old Select Action a The probability of E Express expectations, A ( s, a ) is the advantage function, which means that in the state s Next, select Action a The additional income ε Represents the trimming range, and clip() represents the trimming function.

[0039] A second aspect of an embodiment of the present invention provides a smart clothing industry production control system based on data analysis, including:

[0040] The first unit is used to set up an intelligent data collection platform, based on the data collection agent module in the intelligent data collection platform, combined with the real-time transmission technology of data streams to collect heterogeneous data in clothing production equipment, supply chain management systems and sales terminal systems, for the heterogeneous data, through the rule-based data cleaning algorithm and the learning-based data cleaning algorithm to identify missing values ​​and outliers, clean and convert them into vectorized representations, extract common features through the multi-view data fusion algorithm based on tensor decomposition, generate a comprehensive data set in combination with the vectorized representation, extract multimodal feature representations through the multimodal feature representation learning algorithm and build a digital twin model of clothing production;

[0041] The second unit is used to define the core concepts, attributes and relationships in the field of clothing production through the ontology knowledge modeling method, extract data based on the comprehensive data set, generate triples and construct a clothing production knowledge graph, construct a clothing production data analysis model based on the graph attention network algorithm, add the clothing production knowledge graph to the clothing production data analysis model, adaptively extract key features and determine implicit association rules through the attention mechanism, generate a cross-modal knowledge graph, combine the prior knowledge corresponding to the clothing production knowledge graph, identify clothing production control rules through the knowledge reasoning algorithm, and solve the clothing production control parameters in combination with the multi-objective optimization algorithm;

[0042] The third unit is used to build a decision evaluation model based on the reinforcement learning algorithm and set a reward function, parse the clothing production control parameters into production control decisions, evaluate the production control decisions through the decision evaluation model, and dynamically correct them based on the evaluation results to obtain optimized production control decisions. The optimized production control decisions are sent to the equipment terminals in the clothing production workshop and adaptively optimize clothing production in combination with task requirements, collect real-time production data in the production process and update the clothing production digital twin model and the clothing production knowledge graph, and perform periodic optimization until the production task is completed.

[0043] According to a third aspect of the embodiments of the present invention,

[0044] An electronic device is provided, comprising:

[0045] processor;

[0046] a memory for storing processor-executable instructions;

[0047] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0048] A fourth aspect of the embodiments of the present invention is:

[0049] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0050] In the present invention, the heterogeneous data in clothing production equipment, supply chain management system and sales terminal system are collected through the intelligent data collection platform, the data barriers of design, production, sales and other links are broken, and the data cleaning algorithm combined with rule learning and multi-view data fusion algorithm are adopted to effectively improve the data quality and fusion depth, and provide a high-quality data foundation for subsequent analysis and optimization. Through ontology knowledge modeling, the core concepts, attributes and relationships in the field of clothing production are defined, and the knowledge graph is constructed by extracting triples based on the comprehensive data set, so that the expert knowledge such as production process, quality standard, management procedure can be understood by computer. The formal systematization representation realizes the standardization and structured management of knowledge, provides rich domain knowledge support for intelligent regulation, adaptively extracts the key features of clothing styles, fabrics, processes, equipment and other elements through the attention mechanism, and discovers implicit correlation rules from production practice data to generate cross-modal knowledge graphs, which depict the internal mechanism of clothing production from the multi-dimensional dimensions of data, knowledge and business, and form a systematic and correlated understanding of the overall production. In summary, the present invention constructs a set of clothing production intelligent regulation and optimization system with deep integration of data, knowledge and intelligence, which comprehensively improves the digitalization, networking and intelligence level of clothing production. It has significant effects on improving the quality of clothing products, improving production efficiency, reducing operating costs, and enhancing market responsiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 A schematic diagram of a process of a smart clothing industry production control method based on data analysis according to an embodiment of the present invention;

[0052] Figure 2 It is a structural schematic diagram of an intelligent clothing industry production control system based on data analysis according to an embodiment of the present invention. DETAILED DESCRIPTION

[0053] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0054] The technical solution of the present invention is described in detail with specific embodiments below. The following specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described in detail in some embodiments.

[0055] Figure 1 FIG. 1 is a flow chart of a method for production control in the intelligent clothing industry based on data analysis according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0056] S1. Set up an intelligent data collection platform, based on the data collection agent module in the intelligent data collection platform, combine the data stream real-time transmission technology to collect heterogeneous data in clothing production equipment, supply chain management system and sales terminal system, for the heterogeneous data, use rule-based data cleaning algorithm and learning-based data cleaning algorithm to identify missing values ​​and outliers, clean and convert them into vectorized representation, extract common features through multi-view data fusion algorithm based on tensor decomposition, generate a comprehensive data set in combination with the vectorized representation, extract multimodal feature representation through multimodal feature representation learning algorithm and build a digital twin model of clothing production;

[0057] The data acquisition agent module is a system component used to automatically collect and pre-process data from multiple data sources (including sensors, databases, network services, etc.). The heterogeneous data refers to data from different data sources. The rule-based data cleaning algorithm relies on a predefined set of rules to identify and correct errors or anomalies in the data. The learning-based data cleaning algorithm uses a machine learning model to automatically discover anomalies and errors in the data and correct them. The multi-perspective data fusion algorithm based on tensor decomposition is used to integrate multi-source data from different perspectives (such as time, space, category, etc.). The common features refer to similar or identical attributes, characteristics or behavior patterns that are prevalent in different things, phenomena or systems.

[0058] In an optional embodiment,

[0059] An intelligent data collection platform is set up, and based on the data collection agent module in the intelligent data collection platform, heterogeneous data in clothing production equipment, supply chain management system and sales terminal system are collected in combination with data stream real-time transmission technology. For the heterogeneous data, missing values ​​and abnormal values ​​are identified through rule-based data cleaning algorithms and learning-based data cleaning algorithms, and the data are cleaned and converted into vectorized representations. Common features are extracted through a multi-view data fusion algorithm based on tensor decomposition, and a comprehensive data set is generated in combination with the vectorized representation. The multi-modal feature representation is extracted through a multi-modal feature representation learning algorithm and a digital twin model of clothing production is constructed, including:

[0060] An intelligent data collection platform is set up based on a microservice architecture, wherein the intelligent data collection platform includes a data collection agent module, a message middleware, a data cleaning module and a digital twin engine. The data collection agent module adopts a sidecar mode and is deployed in a containerized manner with the data interfaces of clothing production equipment, supply chain management system and sales terminal system. For the data in the clothing production equipment, equipment parameters and status data are periodically collected. For non-standard private protocols, the edge computing gateway is combined to generate standard equipment data through protocol parsing and data conversion and send it to the message middleware. For the supply chain management system, it is connected through a descriptive status interface, and purchase orders, inventory levels and logistics status are regularly pulled as supply chain data. The supply chain data is extracted into a pre-set data mart and sent to the message middleware. For the sales terminal system, the sales quantity of the e-commerce platform is regularly obtained through data crawling technology. For unstructured user comment data, product attribute words and sentiment polarity are extracted in combination with sentiment analysis and topic mining technology and sent to the message middleware. The data in the message middleware is combined to obtain the heterogeneous data.

[0061] For the structured data in the heterogeneous data, the integrity and conflict of the data are verified by a data quality verification function based on a distributed structured query language engine, and non-compliant data is identified and removed in combination with business rules. For unstructured data, a data preprocessing algorithm based on a distributed machine learning library is used to map words in text data to a low-dimensional vector space using a word embedding algorithm, generate word vectors and perform cleaning, and multi-scale textures and edges in image data are extracted through a convolutional neural network to generate image feature vectors, which are cleaned according to a pre-set quality threshold. For time series data, a symbolic aggregation approximation algorithm is selected for segmented aggregation forcing and a symbolic mapping algorithm is used to map the time series data into a symbolic sequence, reduce the data dimension and extract the morphological features of the time series, and perform outlier identification and missing value completion based on the morphological features;

[0062] During the cleaning process, the thresholds of missing values ​​and outliers are adjusted according to data quality through adaptive threshold strategies. For data that cannot be judged during the cleaning process, the most representative data is selected for annotation through artificial intelligence technology. The cleaning is repeated and the cleaned heterogeneous data is converted into vectorized representation.

[0063] For heterogeneous data represented by vectorization, a high-order tensor is constructed through a multi-perspective data fusion algorithm based on tensor decomposition, wherein each dimension of the high-order tensor corresponds to a data perspective, and the high-order tensor is low-rank approximated through the Tucker decomposition algorithm to obtain a core tensor and multiple factor matrices, and domain ontology knowledge is applied as constraints to the core tensor and the factor matrix. A heterogeneous information network is constructed through a graph-based data fusion strategy, and feature vectors of different perspectives are used as nodes, and the correlation between features is used as edges. The weight of the edge is determined according to the semantic similarity, and the low-dimensional embedding representation of the network nodes is learned by combining the random walk network embedding algorithm and mapped to the same semantic space to obtain common features and construct the comprehensive data set based on the common features, and the multimodal feature representation in the comprehensive data set is extracted through the multimodal feature representation learning algorithm, and a virtual simulation environment of a clothing production workshop is constructed in combination with the digital twin engine, and the entity is abstracted into a three-dimensional model in combination with a pre-set three-dimensional collection library to obtain the clothing production digital twin model.

[0064] The message middleware is a software system mainly used to manage the transmission of messages in a distributed system. The sidecar mode is a microservice architecture mode. An auxiliary service is attached to the main application service as its "sidecar". The containerized deployment is a way of encapsulating applications and their dependencies using container technology. The edge computing gateway is a device located at the edge of the network, responsible for performing computing tasks close to the data source to reduce data transmission delays and network bandwidth consumption. The representational state interface is an architectural style for building network applications. The data mart is a data storage system for specific business areas or departments. The topic mining technology is a natural language processing method for automatically extracting potential topics from a large number of documents. The sentiment polarity refers to the sentiment tendency expressed in the text, which is usually divided into three polarities: positive, negative and neutral. The data quality verification function based on the distributed structured query language engine is used to A function that verifies quality attributes such as data integrity, accuracy, and consistency in a large-scale distributed database. The word embedding algorithm is a technology in natural language processing, which is used to represent words as low-dimensional vectors so that semantically similar words are closer in the vector space. The symbolic aggregation approximation algorithm is a method for dimensionality reduction and representation of time series data. The segmented aggregation approximation is a dimensionality reduction method for time series data, which simplifies the representation of the original sequence by dividing the time series into multiple equal-length sub-segments and calculating the average value of each sub-segment. The low-rank approximation is a matrix decomposition technology used to approximate the original matrix as the product of two or more low-rank matrices. The core tensor is a key element in tensor decomposition, which represents the compact representation of the tensor after dimensionality reduction in multiple dimensions. The factor matrix is ​​a matrix used to decompose the dimensions of the tensor in tensor decomposition. The network embedding algorithm is a technology that represents nodes in a graph or network structure as low-dimensional vectors.

[0065] Build a monitoring system to collect various indicators of microservices in real time, such as request latency and error rate. When the indicators are abnormal, alarms are automatically triggered to notify operation and maintenance personnel to handle them in time. At the same time, use link tracking tools (such as Silk) to realize the visualization of microservice call chains and performance bottleneck analysis;

[0066] The data collection agent adopts the sidecar mode and is deployed in a containerized manner together with the data interfaces of clothing production equipment, supply chain management system, and sales terminal system, decoupling the collection agent from the business system and allowing it to run and upgrade independently. For clothing production equipment, the collection agent uses the IoT protocol to communicate with the equipment, and periodically (e.g., every 5 seconds) collects equipment parameters (e.g., speed, temperature) and status (e.g., operation, fault) data. For non-standard private protocols, an edge gateway (e.g., Raspberry Pi) is deployed on the device side, and private protocol data is converted into standard IoT messages through protocol adaptation and published to the IoT protocol server. For supply chain management systems, enterprise resource planning (ERP) systems such as SAP are generally used, and the collection agent communicates with ERP through a representational state transfer interface. System docking, regularly (e.g., every hour) pull purchase orders, inventory levels, logistics status and other data, extract, clean and transform the pulled data through ETL tools, and finally store it in a structured form in the data mart. For the sales terminal system, the collection agent regularly (e.g., every day) crawls the sales data of e-commerce platforms (such as Tmall and JD.com) through web crawler technology, including sales volume, price, comments, etc. For unstructured user comment texts, use sentiment analysis models to determine sentiment polarity (positive, negative), and use topic models to extract key product attributes involved in the comments (such as fabrics, styles). The collection agent sends equipment data, ERP data, sales data, etc. to the message middleware to provide a data source for subsequent data cleaning and fusion;

[0067] For structured data, the structured query language module of the distributed computing engine is used to verify data quality, such as integrity constraints (such as the order amount cannot be empty), value range constraints (such as the order status can only be enumerated values ​​such as "paid" and "shipped"), foreign key constraints (such as the product number in the order must exist in the product table), etc., to filter out dirty data that does not meet the constraints. For unstructured text data, first use the Chinese word segmentation tool to divide the long text into word sequences, then use the word embedding model to map the words into low-dimensional real number vectors, convert the word sequence into a word vector sequence, and use a sequence cleaning algorithm (such as removing stop words and stemming) to reduce the noise of the word vector sequence, and finally obtain clean text features. For image data, use a deep learning model to extract multi-scale texture features of the image (such as scale-invariant feature transform SIFT) and edge features, use similarity functions (such as cosine similarity) to calculate the similarity between image feature vectors, and filter out images with similarity below a threshold (such as 0.5) as noise images. For time series data, first use the symbolic aggregation approximation SAX algorithm to convert the real number sequence into a symbolic sequence to reduce the data dimension, and use the sliding window algorithm to extract the morphological features of the symbolic sequence, such as linearity, periodicity, mutation, etc. Based on the morphological features, use the Tukey algorithm to detect outliers in the time series, and use interpolation algorithms (such as linear interpolation) to fill in missing values. In the data cleaning process, use active learning algorithms to select the most representative and most uncertain small amount of data, let domain experts perform manual annotation, continuously iterate and update the cleaning rules, and systematically learn expert feedback to converge to the optimal cleaning strategy to achieve incremental improvement of cleaning quality;

[0068] A multi-view data fusion method based on tensor decomposition is adopted. Each data view (such as sales, comments, and pictures) is used as a dimension of the tensor to form a high-order tensor. The high-order tensor is decomposed into a core tensor and multiple factor matrices using the Tucker decomposition algorithm. The domain ontology knowledge (such as two products belong to the same category, so their comments should be similar) is added as a constraint to the objective function of the tensor decomposition, so that the decomposition result conforms to the domain knowledge. A graph-based data fusion method is adopted. The features in different data views are used as nodes of a heterogeneous information network, and the association between features (such as two products purchased by the same user) is used as an edge. The embedding function of nodes and edges is designed so that semantically similar nodes and edges are embedded into similar low-dimensional vector spaces. A network embedding algorithm based on random walks (such as DeepWalk) is used to learn the embedding representation of nodes, and nodes of different modalities are mapped to the same semantic space to achieve feature fusion, obtain the common features, and perform data enhancement on the fused data. Use unsupervised generative models such as variational autoencoders to learn the essential distribution of data, generate synthetic data similar to real data, add the synthetic data to the original data set, expand the data volume, and improve the generalization of the model;

[0069] Select a 3D real-time rendering engine such as Unity3D, build a virtual simulation environment for a clothing production workshop, and construct its 3D geometric model in the virtual space according to the computer-aided design drawings of the workshop, including factory buildings, equipment, products, workers, etc. Map the fused data into the attributes of the virtual object, and abstract the entity into a 3D model in combination with the pre-set 3D collection library to obtain the clothing production digital twin model.

[0070] In this embodiment, containerized deployment improves the portability and environmental consistency of microservices, and uses container orchestration to achieve elastic scaling and rapid recovery of microservices. The regular collection of sales data, especially sentiment analysis and attribute extraction of user review data, provides data support for product improvement and marketing strategy optimization. A small amount of manual annotation and active learning mechanism are introduced, and expert feedback is continuously learned to optimize data cleaning rules, thereby achieving continuous improvement in data cleaning quality at a relatively low labor cost. Data enhancement technology is adopted, and a large amount of realistic synthetic data is generated using generative models, which effectively expands the scale and diversity of the original data set and improves the accuracy and generalization of the data-driven model. In summary, this embodiment realizes the interconnection and asset management of data, empowers traditional manufacturing processes with digital twins and artificial intelligence technologies, and creates a new paradigm for the digital transformation of the apparel industry, which is of great significance for promoting quality improvement and efficiency enhancement in the apparel industry.

[0071] In an optional embodiment,

[0072] The high-order tensor is approximated to a low rank using the Tucker decomposition algorithm as shown in the following formula:

[0073] ;

[0074] in, X represents a high-order tensor, G represents the core tensor, × n represents the modular product operation of tensors, Indicates N factor matrix, N represents the total number of factor matrices, n represents the index of the factor matrix, λ represents the regularization parameter for the smoothness constraint, tr represents the trace of the matrix, L n represents the graph Laplacian matrix, represents the factor matrix, ‖‖ F represents the Frobenius norm.

[0075] In this embodiment, by controlling the rank of the core tensor and the factor matrix, the accuracy and complexity of the approximate representation can be flexibly adjusted, and a trade-off is made between information loss and computational efficiency to achieve the fusion of multi-view data. The information of different views is not simply superimposed, but the interaction and transmission of information is realized in a common semantic space, thereby obtaining a comprehensive and accurate data representation. By constructing a similarity graph of factors, the graph Laplacian matrix can measure the manifold structure between factors. In summary, this embodiment realizes the fusion of multi-source heterogeneous data, and reveals the inherent patterns and laws of the data in an interpretable way. It introduces prior domain knowledge to optimize the data representation, which is of great significance to alleviate data sparsity and heterogeneity, improve the effect and application value of data fusion, and support intelligent decision-making based on data and knowledge.

[0076] S2. Define the core concepts, attributes and relationships in the field of clothing production through the ontology knowledge modeling method, extract data based on the comprehensive data set, generate triples and construct a clothing production knowledge graph, construct a clothing production data analysis model based on the graph attention network algorithm, add the clothing production knowledge graph to the clothing production data analysis model, adaptively extract key features and determine implicit association rules through the attention mechanism, generate a cross-modal knowledge graph, combine the prior knowledge corresponding to the clothing production knowledge graph, identify clothing production control rules through the knowledge reasoning algorithm, and solve and obtain clothing production control parameters in combination with the multi-objective optimization algorithm;

[0077] The ontology knowledge modeling method is a structured method for representing, organizing and managing knowledge. The triple is the basic unit in knowledge representation, which consists of three parts: "subject-predicate-object" and is often used to represent the relationship in the knowledge graph. The graph attention network is a deep learning algorithm for graph structured data. The clothing production knowledge graph is a knowledge graph constructed for the clothing production field, which organizes the information of various links, process flows, materials, equipment, etc. in clothing production in the form of ontology and triples. The clothing production data analysis model is used to analyze and process a large amount of data related to clothing production. The implicit association law refers to the hidden association relationship that is difficult to directly observe in the data. The cross-modal knowledge graph is a knowledge graph constructed by integrating information from different modalities (such as text, image, video, etc.), allowing connections to be established between data in different modalities, thereby supporting more complex reasoning and query. The knowledge reasoning algorithm is used to perform automatic reasoning in the knowledge graph or ontology to discover new knowledge or verify existing knowledge. The clothing production regulation rules are the rules and standards that guide the operation of each link in the clothing production process.

[0078] In an optional embodiment,

[0079] The core concepts, attributes and relationships in the field of clothing production are defined through the ontology knowledge modeling method, data extraction is performed based on the comprehensive data set, triples are generated and a clothing production knowledge graph is constructed, a clothing production data analysis model is constructed based on the graph attention network algorithm, the clothing production knowledge graph is added to the clothing production data analysis model, key features are adaptively extracted and implicit association rules are determined through the attention mechanism, a cross-modal knowledge graph is generated, and the clothing production control rules are identified through the knowledge reasoning algorithm in combination with the prior knowledge corresponding to the clothing production knowledge graph, and the clothing production control parameters are solved in combination with the multi-objective optimization algorithm, including:

[0080] Based on expert interviews and literature research, the core concepts, attributes and relationships in the field of clothing production are obtained, and the knowledge in the field of clothing production is formally represented. Based on the ontology knowledge modeling method, the hierarchical structure is determined, and the superordinate concepts and subordinate concepts are divided. The acquired knowledge is integrated and the production task ontology is defined. The production process is abstracted into a directed acyclic graph, in which nodes represent production tasks and edges represent dependencies between tasks. The semantic mapping relationship between different ontologies is identified through an ontology matching algorithm based on semantic description, and knowledge is fused based on the semantic mapping relationship. The ontology alignment algorithm based on logical reasoning is combined to eliminate semantic objections and perform ontology reasoning and expansion to obtain an ontology knowledge base.

[0081] The comprehensive data set is semantically annotated and knowledge extracted, entities, attributes and relationships are extracted to generate triples, the triples are imported into a preset graph database, attribute indexes of vertices and edges are constructed, the clothing production knowledge graph is obtained, a clothing production data analysis model is constructed based on a graph attention network algorithm, the clothing production knowledge graph is converted into a graph embedding representation and added to the clothing production data analysis model, the semantic representation of vertices in the clothing production knowledge graph is learned through a graph attention network, neighbor information is aggregated through attention weights and its own feature representation is updated, the importance weights of elements in the clothing production knowledge graph are calculated based on preset attention functions of different granularities and the key features are weighted to generate the key features, the key features are interpreted, the frequent co-occurrence patterns and sequence patterns between production factors are determined in combination with an association rule mining algorithm, the semantic association structure is identified in combination with a graph pattern matching technology, the implicit association law is obtained, and the cross-modal knowledge graph is constructed based on the implicit association law and the clothing production knowledge graph;

[0082] Prior knowledge corresponding to the clothing production knowledge graph is obtained, the prior knowledge is integrated with the cross-modal knowledge graph, and the cross-modal knowledge graph integrated with the prior knowledge is logically inferred according to the existing production rules through a knowledge reasoning algorithm to obtain the clothing production control rules. Based on a multi-objective optimization algorithm, an objective function is constructed with the goals of maximizing production efficiency and minimizing quality loss. The objective function is solved based on a genetic algorithm to generate a Pareto optimal solution set and perform optimization search through taboo search to obtain the clothing production control parameters.

[0083] The directed acyclic graph is a graph structure in which the edges have directions and there is no path from a node back to itself. The ontology matching algorithm is used to find similar or equivalent concepts, attributes and relationships between different ontologies. The semantic mapping relationship refers to the semantic correspondence established between different concepts or data models. The ontology alignment algorithm based on logical reasoning uses logical reasoning technology to align different ontologies. By analyzing the logical relationships and constraints in the ontology, the semantic consistency and logical correctness of the alignment results can be ensured. The semantic objection refers to the difference in understanding of the same concept or relationship between different systems, data sources or users. The graph embedding representation is to embed the concepts in the graph structure into The technology of mapping nodes or subgraphs into low-dimensional vector space, the neighbor information refers to other nodes and their attributes directly connected to a node in the graph, the attention function is a mechanism used in neural networks to selectively focus on specific parts of the input data, the interpretable analysis refers to the explanation and understanding of the model output results, especially in complex machine learning models such as deep learning or graph neural networks, the association rule mining algorithm is used to discover frequently occurring patterns and rules from large amounts of data, the graph pattern matching technology is used to find subgraphs that match a given pattern graph in a large-scale graph structure, and the logical inference is the process of deriving new knowledge based on logical rules and known facts.

[0084] Through expert interviews and literature research, we obtain the core concepts, attributes and relationships in the field of clothing production, such as clothing style, fabric, process, equipment and other concepts, style color, fabric composition and other attributes, style and fabric matching relationship, etc. We use ontology knowledge modeling methods (such as RDF, OWL) to formally represent the acquired knowledge and clearly define the hierarchical structure of concepts. For example, clothing categories are divided into upper-level concepts such as tops, pants, and skirts, and shirts and T-shirts are divided into lower-level concepts of tops. We integrate multi-source heterogeneous clothing production knowledge and define a unified production task ontology. For example, production links such as cutting, sewing, and ironing are abstracted into task concepts in the ontology, and production tasks are divided into task concepts in the ontology. The process is abstracted as a directed acyclic graph, where nodes represent production tasks and edges represent dependencies between tasks, such as sewing tasks depend on cutting tasks. An ontology matching algorithm based on semantic description (such as semantic similarity calculation based on WordNet) is used to identify semantic mapping relationships between different ontologies, such as the mapping between the concept of "cotton" in the fabric ontology and the concept of "cotton" in the supplier ontology. Ontology knowledge is integrated based on semantic mapping relationships, and an ontology alignment algorithm based on logical reasoning (such as the Tableau algorithm based on description logic) is used to eliminate concept ambiguity. Knowledge reasoning and expansion are performed through ontology axioms to build an ontology knowledge base in the field of clothing production.

[0085] Semantically annotate multi-source heterogeneous clothing production data (such as work orders, materials, equipment parameters, etc.), map unstructured text into structured triples, such as <Dress A, fabric used, chiffon X>, use knowledge extraction technology (such as conditional random field CRF) to extract entity, attribute, and relationship triples from structured and semi-structured data, such as extracting <chiffon X, supplier, supplier B> from the supplier database, import the extracted triples into a graph database (such as Neo4j), and construct a knowledge graph with entities as vertices and relationships as edges. Each vertex and edge has attribute information, such as the color and thickness attributes of chiffon fabric X. Build attribute indexes for vertices and edges in the graph to improve the retrieval and query efficiency of the graph. Finally, a clothing production knowledge graph that integrates multi-dimensional information such as clothing products, materials, processes, and suppliers is obtained;

[0086] Use graph attention network algorithms (such as GAT) to build a clothing production data analysis model, use knowledge graphs as input to the model, learn low-dimensional semantic representations of vertices, dynamically assign different weights to the neighbors of vertices through the attention mechanism, aggregate neighbor information to update the features of the vertex itself, such as when generating chiffon fabric features, the chemical property edges connected to it should be given higher weights, design attention functions of different granularities, extract important features of vertices from local and global perspectives, such as the first-order neighbors of materials and suppliers reflect the physical properties of materials, and reveal the correlation features of upstream and downstream supply chains through multi-order neighbors. Based on the weighted sum of attention weights, generate semantic feature representations of key elements such as clothing styles, fabrics, and craftsmanship in the knowledge graph, use interpretable analysis technology to analyze the importance of different features to downstream production tasks, and identify key factors affecting clothing quality and delivery time;

[0087] Use frequent item mining algorithms (such as Apriori) and sequential pattern mining algorithms (such as GSP) to discover the frequent co-occurrence and temporal rules of different production factors from the clothing production knowledge graph, such as the frequent co-purchase of a certain fabric and a certain auxiliary material. Use graph pattern matching technology to identify the association structure in the knowledge graph, such as the chain relationship of supplier-fabric-clothing style-sales store, and discover the dependency relationship between each link of the supply chain. Combine frequent co-occurrence patterns, sequence patterns, and graph patterns to summarize the implicit association rules of the production process, such as the best combination of the supplier of fabrics and auxiliary materials and processing technology for a certain hot-selling clothing. The prior knowledge of clothing production obtained from the ontology knowledge base, such as material applicability principles and style design specifications, is integrated with the implicit association rules obtained from data-driven mining. The knowledge reasoning algorithm is used on the integrated knowledge graph to derive new production decision-making knowledge based on empirical rules in material selection, process optimization, etc., and the prior knowledge is integrated with data association rules to build a cross-modal knowledge graph of clothing production that spans materials, processes, supply chains, and sales.

[0088] Based on the cross-modal knowledge graph, a multi-objective optimization algorithm (such as NSGA-II) is used to model and solve the production scheduling problem. The optimization goals are to maximize production efficiency and minimize quality loss, and a multi-objective function is constructed. Each optimization variable corresponds to a production factor in the knowledge graph, such as fabric selection, cutting process, etc. The value of the variable is the entity or attribute value in the knowledge graph. A genetic algorithm is used to solve the multi-objective function. Based on the association rules in the knowledge graph (such as the adaptability of fabrics and styles), crossover and mutation operators are designed to make the optimization process meet the production constraints and generate multiple Pareto optimal solutions, representing different trade-offs between production efficiency and quality. The taboo search algorithm is used to perform local search on the Pareto solution set to obtain the optimal combination of production control parameters, and the optimized parameter combination is converted into control rules for production personnel, such as the optimal production plan for dress A: select chiffon fabric X, cutting size tolerance ±2mm, and sewing machine speed 800r / min.

[0089] In this embodiment, the formal representation of knowledge in the field of clothing production is realized through ontology construction, the definitions of conceptual terms and attribute relationships are unified, knowledge islands are eliminated, and the foundation is laid for cross-departmental and cross-enterprise knowledge sharing. The multi-dimensional associations between materials, processes, styles, and supply chains are established through knowledge graph technology, which not only mines explicit data dependencies, but also discovers implicit business rules, such as the matching principles of fabric selection and style design, and the timing constraints of production processes, forming a systematic understanding of the entire process of clothing production. Through knowledge reasoning technology, ontology rules, association patterns and production data are combined in real time, and material selection and process parameters at the production site are timely reasoned and optimized, which improves the timeliness of production decisions. The reasoning rules that integrate domain knowledge are interpretable and ensure the credibility of the reasoning results. Through multi-objective optimization methods, production efficiency and product quality are reasonably weighed, and Pareto optimal control parameters are generated, achieving a dynamic balance of production goals and improving the scientific nature of production management. In summary, this embodiment realizes the full-process knowledge empowerment from material selection, process optimization to supply chain management, forming a set of effective intelligent optimization solutions for clothing production, which is of great significance to improving the competitiveness of the industry and promoting the intelligent transformation of traditional manufacturing.

[0090] In an optional embodiment,

[0091] Based on the multi-objective optimization algorithm, the objective function is constructed with the goal of maximizing production efficiency and minimizing quality loss as shown in the following formula:

[0092] ;

[0093] in, x i Indicates production tasks i The processing time, ti Indicates the task i Standard processing time, c i Indicates the task i The mass loss coefficient, μ i Indicates production tasks i The mass loss rate, m Indicates the total number of tasks, represents the actual value of mass loss, Indicates the actual value of production efficiency.

[0094] In this embodiment, by using normalization processing, indicators of different magnitudes can be unified to the same scale, so that the objective function can fairly evaluate production efficiency and quality loss, and can effectively avoid conflicts between production efficiency and quality loss, ensuring that product quality is not sacrificed while pursuing high efficiency. Both production efficiency and quality loss have undergone normalization processing of maximum and minimum values, and have adaptive capabilities. They can automatically adjust the optimization direction and intensity according to the actual situation of the current task. By comprehensively evaluating multiple indicators, finding a relatively balanced optimal solution is helpful to achieve more comprehensive optimization effects in complex production environments. In summary, this embodiment can achieve effective control and optimization of production efficiency and quality loss, thereby improving the efficiency of the overall production system and achieving a comprehensive balance of production goals under the framework of multi-objective optimization.

[0095] S3. Construct a decision evaluation model based on the reinforcement learning algorithm and set a reward function, parse the clothing production control parameters into production control decisions, evaluate the production control decisions through the decision evaluation model, and dynamically correct them based on the evaluation results to obtain optimized production control decisions. Send the optimized production control decisions to the equipment terminals in the clothing production workshop and adaptively optimize clothing production in combination with task requirements, collect real-time production data during the production process, update the clothing production digital twin model and the clothing production knowledge graph, and perform periodic optimization until the production task is completed.

[0096] The clothing production control parameters refer to key variables that need to be monitored and adjusted during the clothing production process, such as temperature, humidity, sewing speed, pressure, tension, etc. The production control decision refers to the operation adjustment or strategy selection made during the clothing production process based on real-time monitoring data and preset production goals. The equipment terminal is the operating interface of various mechanical equipment or control systems used in the clothing production line. The real-time production data refers to various data collected in real time through sensors, equipment terminals, etc. during the clothing production process, including production speed, temperature, humidity, equipment status, etc.

[0097] In an optional embodiment,

[0098] A decision evaluation model is constructed based on a reinforcement learning algorithm and a reward function is set. The clothing production control parameters are parsed into production control decisions. The production control decisions are evaluated through the decision evaluation model. Dynamic corrections are made based on the evaluation results to obtain optimized production control decisions. The optimized production control decisions are sent to equipment terminals in the clothing production workshop and clothing production is adaptively optimized in combination with task requirements. Real-time production data in the production process is collected and the clothing production digital twin model and the clothing production knowledge graph are updated. Periodic optimization is performed until the production task is completed, including:

[0099] Based on the reinforcement learning algorithm, according to the characteristics of clothing production regulation and control decisions, the clothing production regulation parameters are used as state variables to construct a multidimensional state space, the clothing production regulation decisions are used as action variables to construct an action space, a multi-objective reward function is set according to production efficiency, quality loss and energy consumption, and the decision evaluation model is constructed based on the multi-dimensional state space, the action space and the multi-objective reward function;

[0100] The clothing production control parameters are parsed into production control decisions and added as input data to the decision evaluation model, the expected cumulative reward corresponding to the current production control decision is calculated through forward propagation, and the expected cumulative reward is compared with a preset reward threshold. If the expected cumulative reward is greater than the reward threshold, the current production control decision is considered to be an optimized production control decision. Otherwise, the production control decision is optimized through a policy gradient algorithm to obtain an optimized production control decision.

[0101] The optimized production control decision is sent to the equipment terminal of the clothing production workshop. The equipment terminal adaptively optimizes the clothing production process according to the received decision instructions and the production task requirements, monitors the clothing production status in real time and dynamically adjusts the process parameters and resource allocation, collects real-time production data under different production states, transmits the real-time production data to the clothing production digital twin model, updates it through data synchronization and mapping, generates simulation data through the updated clothing production digital twin model, and updates the clothing production knowledge graph and the cross-modal knowledge graph;

[0102] An optimization cycle is set and the optimization is performed cyclically. In each optimization cycle, the optimized production control decision is repeatedly executed and generated until the production task is completed or the preset optimization goal is reached.

[0103] The multidimensional state space is a multidimensional vector space used to describe the state of the system in reinforcement learning or optimization algorithms, and each dimension corresponds to a state variable of the system, such as temperature, humidity, equipment status, etc. in clothing production. The action variable refers to the control variable or decision variable that the system can execute during the production control process. The action space is the set of all possible actions that the system can choose in reinforcement learning. The quality loss refers to the loss caused by the product quality not meeting the expected standards due to deviations or errors in the production process. The expected cumulative reward is a key indicator for measuring the quality of a strategy in reinforcement learning, and refers to the total reward expected to be obtained by the system from the current state to execute a sequence of actions under the current strategy. The policy gradient algorithm is a reinforcement learning algorithm that maximizes the expected cumulative reward by directly optimizing the parameters of the strategy.

[0104] Taking clothing production control parameters as state variables, constructing a multidimensional state space. State variables may include fabric type, cutting size, sewing process, template parameters, etc. Taking clothing production control decisions as action variables, constructing an action space. Action variables may include material selection, process sequencing, equipment scheduling, personnel allocation, etc. A multi-objective reward function is set according to production efficiency, quality loss and energy consumption. For example, production efficiency can be represented by output per unit time, quality loss can be represented by defective rate and return rate, and energy consumption can be represented by energy consumption per unit product. Based on the multidimensional state space, action space and multi-objective reward function, a decision evaluation model is constructed;

[0105] The clothing production control parameters are parsed into production control decisions, which are added as input data to the decision evaluation model. The expected cumulative reward corresponding to the current production control decision is calculated through forward propagation, and the expected cumulative reward is compared with the preset reward threshold. If the expected cumulative reward is greater than the reward threshold, the current production control decision is considered to be optimized. Otherwise, the production control decision is optimized through the policy gradient algorithm to obtain the optimized production control decision. The optimization is repeated until the reward threshold is met or the maximum number of iterations is reached. The optimized production control decision is sent to the equipment terminal in the clothing production workshop. The equipment terminal adaptively optimizes the clothing production process according to the received decision instructions and the production task requirements, monitors the clothing production status in real time, including equipment conditions, material consumption, work-in-progress quality, etc., and dynamically adjusts process parameters and resource allocation, such as optimizing sewing parameters, balancing production rhythm, and dynamic production scheduling.

[0106] Collect real-time production data under different production conditions, including processing data, logistics data, quality data, etc., and transmit the real-time production data to the digital twin model of clothing production. Update the digital twin model of clothing production through data synchronization and mapping, generate simulation data through the updated digital twin model, and update the clothing production knowledge graph and cross-modal knowledge graph. Set optimization cycles, such as each shift, daily, weekly, etc., and repeatedly generate new optimized production control decisions in each optimization cycle until the production task is completed or the preset optimization goals are achieved, such as increasing production capacity by 10% or reducing costs by 5%.

[0107] In this embodiment, a real-time closed-loop decision optimization mechanism is formed by constructing a multi-dimensional state space, action space and reward function. The system can dynamically adjust process parameters, equipment and personnel configuration according to changes in production status, so that the production process is always in the optimal state, which improves the flexibility and adaptability of production management. A multi-objective reward function design is adopted to quantify and weight each core indicator, and the idea of ​​Pareto optimality is introduced. Multiple objectives are comprehensively weighed during decision optimization, and balanced optimization of the production process is achieved. The negative impact that may be caused by the optimization of a single indicator is overcome. By constructing a general knowledge system in the field of clothing production and establishing an association mapping between core elements such as products, processes, and equipment, the optimization decision model can be highlighted to a certain extent. Breaking the category restrictions, the knowledge learned from the optimization of one product can be transferred and applied to other similar products, which improves the generalization ability of optimization, and digitally maps and semantically associates the entire production process, so that the control decision can be coordinated from a global perspective, and realizes the coordinated optimization of each link from raw material input to finished product output, and maximizes the overall efficiency. In summary, this embodiment realizes the real-time dynamic optimization of production control, drives knowledge with data, guides decisions with knowledge, and optimizes actions with decisions, forming a two-way iterative upgrade of production practice and knowledge experience, and provides new ideas and methods for the intelligent, refined and flexible management of clothing production, which has important theoretical significance and practical value for enhancing the competitiveness of clothing enterprises and promoting the transformation and upgrading of the industry.

[0108] In an optional embodiment,

[0109] The production control decision is optimized by the policy gradient algorithm as shown in the following formula:

[0110] ;

[0111] in, Indicates in status s Next, the current strategy π θ Select Action a The probability of Indicates in status s Next, the old strategyπ old Select Action a The probability of E Express expectations, A ( s , a ) is the advantage function, which means that in the state s Next, select Action a The additional income ε Represents the trimming range, and clip() represents the trimming function.

[0112] In this embodiment, an optimization method based on policy gradient is adopted, and a clipping operation is used to limit the amplitude of policy updates, which helps to prevent drastic changes in the policy during the update process, thereby maintaining the stability and reliability of the policy. By limiting the range of changes of the new policy relative to the old policy, the policy collapse phenomenon in the training process can be avoided and the convergence of the model can be improved. By introducing the advantage function, the relative advantages of different actions in the current state can be evaluated, which helps the strategy to make full use of existing excellent strategies and continue to explore new potential optimal strategies during the optimization process, so as to find better control decisions in a variety of production environments. By maximizing the expected return to optimize the strategy, the strategy can gradually tend to those actions that can bring higher cumulative returns in the long run, ultimately improving production efficiency and reducing quality losses. In summary, this embodiment can significantly improve the stability, robustness and expected return of the decision, and control the deviation of the policy update to a certain extent, and ultimately achieve effective management and optimization of the production process.

[0113] Figure 2 FIG. 1 is a schematic diagram of the structure of the intelligent clothing industry production control system based on data analysis according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0114] The first unit is used to set up an intelligent data collection platform, based on the data collection agent module in the intelligent data collection platform, combined with the real-time transmission technology of data streams to collect heterogeneous data in clothing production equipment, supply chain management systems and sales terminal systems, for the heterogeneous data, through the rule-based data cleaning algorithm and the learning-based data cleaning algorithm to identify missing values ​​and outliers, clean and convert them into vectorized representations, extract common features through the multi-view data fusion algorithm based on tensor decomposition, generate a comprehensive data set in combination with the vectorized representation, extract multimodal feature representations through the multimodal feature representation learning algorithm and build a digital twin model of clothing production;

[0115] The second unit is used to define the core concepts, attributes and relationships in the field of clothing production through the ontology knowledge modeling method, extract data based on the comprehensive data set, generate triples and construct a clothing production knowledge graph, construct a clothing production data analysis model based on the graph attention network algorithm, add the clothing production knowledge graph to the clothing production data analysis model, adaptively extract key features and determine implicit association rules through the attention mechanism, generate a cross-modal knowledge graph, combine the prior knowledge corresponding to the clothing production knowledge graph, identify clothing production control rules through the knowledge reasoning algorithm, and solve the clothing production control parameters in combination with the multi-objective optimization algorithm;

[0116] The third unit is used to build a decision evaluation model based on the reinforcement learning algorithm and set a reward function, parse the clothing production control parameters into production control decisions, evaluate the production control decisions through the decision evaluation model, and dynamically correct them based on the evaluation results to obtain optimized production control decisions. The optimized production control decisions are sent to the equipment terminals in the clothing production workshop and adaptively optimize clothing production in combination with task requirements, collect real-time production data in the production process and update the clothing production digital twin model and the clothing production knowledge graph, and perform periodic optimization until the production task is completed.

[0117] According to a third aspect of the embodiments of the present invention,

[0118] An electronic device is provided, comprising:

[0119] processor;

[0120] a memory for storing processor-executable instructions;

[0121] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0122] A fourth aspect of the embodiments of the present invention is:

[0123] A computer-readable storage medium is provided, on which computer program instructions are stored. When the computer program instructions are executed by a processor, the aforementioned method is implemented.

[0124] The present invention may be a method, an apparatus, a system and / or a computer program product. The computer program product may include a computer-readable storage medium carrying computer-readable program instructions for executing various aspects of the present invention.

[0125] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or replace some or all of the technical features therein by equivalents. However, these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A production control method for the intelligent clothing industry based on data analysis, characterized in that: include: An intelligent data collection platform is set up, and based on the data collection agent module in the intelligent data collection platform, heterogeneous data in clothing production equipment, supply chain management system and sales terminal system are collected in combination with data stream real-time transmission technology. For the heterogeneous data, missing values ​​and outliers are identified through rule-based data cleaning algorithms and learning-based data cleaning algorithms, and the data are cleaned and converted into vectorized representations. Common features are extracted through a multi-view data fusion algorithm based on tensor decomposition, and a comprehensive data set is generated in combination with the vectorized representation. Multimodal feature representations are extracted through a multimodal feature representation learning algorithm and a digital twin model of clothing production is constructed; The core concepts, attributes and relationships in the field of clothing production are defined through the ontology knowledge modeling method, data extraction is performed based on the comprehensive data set, triples are generated and a clothing production knowledge graph is constructed, a clothing production data analysis model is constructed based on the graph attention network algorithm, the clothing production knowledge graph is added to the clothing production data analysis model, key features are adaptively extracted and implicit association rules are determined through the attention mechanism, a cross-modal knowledge graph is generated, the clothing production control rules are identified through the knowledge reasoning algorithm in combination with the prior knowledge corresponding to the clothing production knowledge graph, and the clothing production control parameters are solved in combination with the multi-objective optimization algorithm; A decision evaluation model is constructed based on a reinforcement learning algorithm and a reward function is set. The clothing production control parameters are parsed into production control decisions. The production control decisions are evaluated through the decision evaluation model. Dynamic corrections are made based on the evaluation results to obtain optimized production control decisions. The optimized production control decisions are sent to the equipment terminals in the clothing production workshop and clothing production is adaptively optimized in combination with task requirements. Real-time production data in the production process is collected and the clothing production digital twin model and the clothing production knowledge graph are updated. Periodic optimization is performed until the production task is completed. An intelligent data collection platform is set up, and based on the data collection agent module in the intelligent data collection platform, heterogeneous data in clothing production equipment, supply chain management system and sales terminal system are collected in combination with data stream real-time transmission technology. For the heterogeneous data, missing values ​​and abnormal values ​​are identified through rule-based data cleaning algorithms and learning-based data cleaning algorithms, and the data are cleaned and converted into vectorized representations. Common features are extracted through a multi-view data fusion algorithm based on tensor decomposition, and a comprehensive data set is generated in combination with the vectorized representation. The multi-modal feature representation is extracted through a multi-modal feature representation learning algorithm and a digital twin model of clothing production is constructed, including: An intelligent data collection platform is set up based on a microservice architecture, wherein the intelligent data collection platform includes a data collection agent module, a message middleware, a data cleaning module and a digital twin engine. The data collection agent module adopts a sidecar mode and is deployed in a containerized manner with the data interfaces of clothing production equipment, supply chain management system and sales terminal system. For the data in the clothing production equipment, equipment parameters and status data are periodically collected. For non-standard private protocols, the edge computing gateway is combined to generate standard equipment data through protocol parsing and data conversion and send it to the message middleware. For the supply chain management system, it is connected through a descriptive status interface, and purchase orders, inventory levels and logistics status are regularly pulled as supply chain data. The supply chain data is extracted into a pre-set data mart and sent to the message middleware. For the sales terminal system, the sales quantity of the e-commerce platform is regularly obtained through data crawling technology. For unstructured user comment data, product attribute words and sentiment polarity are extracted in combination with sentiment analysis and topic mining technology and sent to the message middleware. The data in the message middleware is combined to obtain the heterogeneous data. For the structured data in the heterogeneous data, the integrity and conflict of the data are verified by a data quality verification function based on a distributed structured query language engine, and non-compliant data is identified and removed in combination with business rules. For unstructured data, a data preprocessing algorithm based on a distributed machine learning library is used to map words in text data to a low-dimensional vector space using a word embedding algorithm, generate word vectors and perform cleaning, and multi-scale textures and edges in image data are extracted through a convolutional neural network to generate image feature vectors, which are cleaned according to a pre-set quality threshold. For time series data, a symbolic aggregation approximation algorithm is selected for segmented aggregation forcing and a symbolic mapping algorithm is used to map the time series data into a symbolic sequence, reduce the data dimension and extract the morphological features of the time series, and perform outlier identification and missing value completion based on the morphological features; During the cleaning process, the thresholds of missing values ​​and outliers are adjusted according to data quality through adaptive threshold strategies. For data that cannot be judged during the cleaning process, the most representative data is selected for annotation through artificial intelligence technology. The cleaning is repeated and the cleaned heterogeneous data is converted into vectorized representation. For heterogeneous data represented by vectorization, a high-order tensor is constructed through a multi-view data fusion algorithm based on tensor decomposition, wherein each dimension of the high-order tensor corresponds to a data view. The high-order tensor is low-rank approximated through the Tucker decomposition algorithm to obtain a core tensor and multiple factor matrices. Domain ontology knowledge is applied as a constraint to the core tensor and the factor matrix. A heterogeneous information network is constructed through a graph-based data fusion strategy, feature vectors of different viewpoints are used as nodes, and the correlation between features is used as edges. The weight of the edge is determined according to the semantic similarity. The low-dimensional embedding representation of the network nodes is learned by combining the random walk network embedding algorithm and mapped to the same semantic space, common features are obtained, and the comprehensive data set is constructed based on the common features. The multimodal feature representation in the comprehensive data set is extracted through the multimodal feature representation learning algorithm. In combination with the digital twin engine, a virtual simulation environment of a clothing production workshop is constructed. In combination with a pre-set three-dimensional collection library, the entity is abstracted into a three-dimensional model to obtain the clothing production digital twin model. The core concepts, attributes and relationships in the field of clothing production are defined through the ontology knowledge modeling method, data extraction is performed based on the comprehensive data set, triples are generated and a clothing production knowledge graph is constructed, a clothing production data analysis model is constructed based on the graph attention network algorithm, the clothing production knowledge graph is added to the clothing production data analysis model, key features are adaptively extracted and implicit association rules are determined through the attention mechanism, a cross-modal knowledge graph is generated, and the clothing production control rules are identified through the knowledge reasoning algorithm in combination with the prior knowledge corresponding to the clothing production knowledge graph, and the clothing production control parameters are solved in combination with the multi-objective optimization algorithm, including: Based on expert interviews and literature research, the core concepts, attributes and relationships in the field of clothing production are obtained, and the knowledge in the field of clothing production is formally represented. Based on the ontology knowledge modeling method, the hierarchical structure is determined, and the superordinate concepts and subordinate concepts are divided. The acquired knowledge is integrated and the production task ontology is defined. The production process is abstracted into a directed acyclic graph, in which nodes represent production tasks and edges represent dependencies between tasks. The semantic mapping relationship between different ontologies is identified through an ontology matching algorithm based on semantic description, and knowledge is fused based on the semantic mapping relationship. The ontology alignment algorithm based on logical reasoning is combined to eliminate semantic objections and perform ontology reasoning and expansion to obtain an ontology knowledge base. The comprehensive data set is semantically annotated and knowledge extracted, entities, attributes and relationships are extracted to generate triples, the triples are imported into a preset graph database, attribute indexes of vertices and edges are constructed, the clothing production knowledge graph is obtained, a clothing production data analysis model is constructed based on a graph attention network algorithm, the clothing production knowledge graph is converted into a graph embedding representation and added to the clothing production data analysis model, the semantic representation of vertices in the clothing production knowledge graph is learned through a graph attention network, neighbor information is aggregated through attention weights and its own feature representation is updated, the importance weights of elements in the clothing production knowledge graph are calculated based on preset attention functions of different granularities and the key features are weighted to generate the key features, the key features are interpreted, the frequent co-occurrence patterns and sequence patterns between production factors are determined in combination with an association rule mining algorithm, the semantic association structure is identified in combination with a graph pattern matching technology, the implicit association law is obtained, and the cross-modal knowledge graph is constructed based on the implicit association law and the clothing production knowledge graph; Prior knowledge corresponding to the clothing production knowledge graph is obtained, the prior knowledge is integrated with the cross-modal knowledge graph, and the cross-modal knowledge graph integrated with the prior knowledge is logically inferred according to the existing production rules through a knowledge reasoning algorithm to obtain the clothing production control rules. Based on a multi-objective optimization algorithm, an objective function is constructed with the goals of maximizing production efficiency and minimizing quality loss. The objective function is solved based on a genetic algorithm to generate a Pareto optimal solution set and perform optimization search through taboo search to obtain the clothing production control parameters.

2. The method according to claim 1, characterized in that The high-order tensor is approximated to a low rank using the Tucker decomposition algorithm as shown in the following formula: ； in, X represents a high-order tensor, G represents the core tensor, × N represents the modular product operation of tensors, Indicates N factor matrix, N represents the total number of factor matrices, n represents the index of the factor matrix, λ represents the regularization parameter for the smoothness constraint, tr represents the trace of the matrix, L n represents the graph Laplacian matrix, represents the factor matrix, ‖‖ F represents the Frobenius norm.

3. The method according to claim 1, characterized in that Based on the multi-objective optimization algorithm, the objective function is constructed with the goal of maximizing production efficiency and minimizing quality loss as shown in the following formula: ； in, x i Indicates production tasks i The processing time, t i Indicates the task i Standard processing time, c i Indicates the task i The mass loss coefficient, μ i Indicates production tasks i The mass loss rate, m Indicates the total number of tasks, represents the actual value of mass loss, Indicates the actual value of production efficiency.

4. The method according to claim 1, characterized in that A decision evaluation model is constructed based on a reinforcement learning algorithm and a reward function is set. The clothing production control parameters are parsed into production control decisions. The production control decisions are evaluated through the decision evaluation model. Dynamic corrections are made based on the evaluation results to obtain optimized production control decisions. The optimized production control decisions are sent to equipment terminals in the clothing production workshop and clothing production is adaptively optimized in combination with task requirements. Real-time production data in the production process is collected and the clothing production digital twin model and the clothing production knowledge graph are updated. Periodic optimization is performed until the production task is completed, including: Based on the reinforcement learning algorithm, according to the characteristics of clothing production regulation and control decisions, the clothing production regulation parameters are used as state variables to construct a multidimensional state space, the clothing production regulation decisions are used as action variables to construct an action space, a multi-objective reward function is set according to production efficiency, quality loss and energy consumption, and the decision evaluation model is constructed based on the multi-dimensional state space, the action space and the multi-objective reward function; The clothing production control parameters are parsed into production control decisions and added as input data to the decision evaluation model, the expected cumulative reward corresponding to the current production control decision is calculated through forward propagation, and the expected cumulative reward is compared with a preset reward threshold. If the expected cumulative reward is greater than the reward threshold, the current production control decision is considered to be an optimized production control decision. Otherwise, the production control decision is optimized through a policy gradient algorithm to obtain an optimized production control decision. The optimized production control decision is sent to the equipment terminal of the clothing production workshop. The equipment terminal adaptively optimizes the clothing production process according to the received decision instructions and the production task requirements, monitors the clothing production status in real time and dynamically adjusts the process parameters and resource allocation, collects real-time production data under different production states, transmits the real-time production data to the clothing production digital twin model, updates it through data synchronization and mapping, generates simulation data through the updated clothing production digital twin model, and updates the clothing production knowledge graph and the cross-modal knowledge graph; An optimization cycle is set and the optimization is performed cyclically. In each optimization cycle, the optimized production control decision is repeatedly executed and generated until the production task is completed or the preset optimization goal is reached.

5. The method according to claim 4, characterized in that The production control decision is optimized by the policy gradient algorithm as shown in the following formula: ； in, Indicates in status s Next, the current strategy π θ Select Action a The probability of Indicates in status s Next, the old strategy π old Select Action a The probability of E Express expectations, A ( s , a ) is the advantage function, which means that in the state s Next, select Action a The additional income ε Represents the trimming range, and clip() represents the trimming function.

6. An intelligent clothing industry production control system based on data analysis, used to implement the method described in any one of claims 1 to 5, characterized in that: include: The first unit is used to set up an intelligent data collection platform, based on the data collection agent module in the intelligent data collection platform, combined with the real-time transmission technology of data streams to collect heterogeneous data in clothing production equipment, supply chain management systems and sales terminal systems, for the heterogeneous data, through the rule-based data cleaning algorithm and the learning-based data cleaning algorithm to identify missing values ​​and outliers, clean and convert them into vectorized representations, extract common features through the multi-view data fusion algorithm based on tensor decomposition, generate a comprehensive data set in combination with the vectorized representation, extract multimodal feature representations through the multimodal feature representation learning algorithm and build a digital twin model of clothing production; The second unit is used to define the core concepts, attributes and relationships in the field of clothing production through the ontology knowledge modeling method, extract data based on the comprehensive data set, generate triples and construct a clothing production knowledge graph, construct a clothing production data analysis model based on the graph attention network algorithm, add the clothing production knowledge graph to the clothing production data analysis model, adaptively extract key features and determine implicit association rules through the attention mechanism, generate a cross-modal knowledge graph, combine the prior knowledge corresponding to the clothing production knowledge graph, identify clothing production control rules through the knowledge reasoning algorithm, and solve the clothing production control parameters in combination with the multi-objective optimization algorithm; The third unit is used to build a decision evaluation model based on the reinforcement learning algorithm and set a reward function, parse the clothing production control parameters into production control decisions, evaluate the production control decisions through the decision evaluation model, and dynamically correct them based on the evaluation results to obtain optimized production control decisions. The optimized production control decisions are sent to the equipment terminals in the clothing production workshop and adaptively optimize clothing production in combination with task requirements, collect real-time production data in the production process and update the clothing production digital twin model and the clothing production knowledge graph, and perform periodic optimization until the production task is completed.

7. An electronic device, characterized in that: include: processor; a memory for storing processor-executable instructions; The processor is configured to call the instructions stored in the memory to execute the method described in any one of claims 1 to 5.

8. A computer-readable storage medium having computer program instructions stored thereon, characterized in that: When the computer program instructions are executed by a processor, the method according to any one of claims 1 to 5 is implemented.

Images