Production tracing and transaction collaboration method and system of MES system

Abstract

The present invention provides a production traceability and transaction collaboration method and system for an MES system, which relates to the technical field of production traceability, including using graph database technology to construct a production process knowledge graph, mapping the fused multi-source heterogeneous data to the entities and relationships of the knowledge graph; based on a time series graph algorithm, establishing a dynamic time series association in the knowledge graph to form a traceable production process chain; based on a knowledge graph and anomaly detection model, constructing a transaction trigger mechanism to automatically identify production transactions that need to be collaboratively processed and obtain transaction collaboration records; using a reinforcement learning algorithm to establish a multi-agent collaborative decision-making model, which can adaptively allocate tasks and resources according to the current production status and historical experience; through a sequence-to-sequence model enhanced by an attention mechanism, converting unstructured collaboration requirements into executable workflows; and automatically generating a multi-dimensional production analysis report based on the traceable production process chain and transaction collaboration records.

Description

Technical Field

[0001] The present invention relates to production tracing technology, and in particular to a production tracing and transaction collaboration method and system of a MES system. Background Art

[0002] With the rapid development of industry, the manufacturing industry has put forward higher requirements for refined management and real-time traceability of the production process. Traditional manufacturing execution systems (MES) face huge challenges in processing heterogeneous data in complex production environments, achieving comprehensive traceability and collaborative decision-making.

[0003] The existing MES system has the following main defects:

[0004] Insufficient data integration and processing capabilities: Traditional MES systems often adopt a centralized data processing architecture, which makes it difficult to effectively process massive heterogeneous data from different sources. In the process of data collection, cleaning and fusion, problems such as data islands and inconsistent information often occur, affecting the real-time monitoring and accurate traceability of the production process.

[0005] The integrity and credibility of production process traceability are not high: the existing system usually uses relational databases to store production data, which makes it difficult to effectively express and analyze complex production process relationships. At the same time, there is a lack of reliable anti-tampering mechanisms, and the authenticity and integrity of traceability data cannot be guaranteed, which poses potential risks in quality traceability and responsibility identification.

[0006] Limited collaborative decision-making capabilities: Traditional MES systems rely on preset rules and manual intervention in abnormal event handling and resource scheduling, and lack intelligent collaborative decision-making mechanisms. Faced with a complex and changing production environment, it is difficult to achieve fast and accurate transaction collaboration and resource optimization, which affects production efficiency and emergency response capabilities.

[0007] Therefore, there is an urgent need for a new MES system that can effectively integrate multi-source heterogeneous data, achieve comprehensive and reliable traceability, and have intelligent collaborative decision-making capabilities to meet the needs of modern manufacturing for efficient, flexible and intelligent production management. Summary of the invention

[0008] The embodiments of the present invention provide a production tracing and transaction collaboration method and system for an MES system, which can solve the problems in the prior art.

[0009] According to a first aspect of the embodiments of the present invention,

[0010] Provide production traceability and transaction collaboration methods for the MES system, including:

[0011] Deploy a distributed data collection network to collect heterogeneous data in real time from production equipment, environmental sensors, human-computer interaction terminals and enterprise information systems; classify and annotate the collected heterogeneous data and establish a unified data model architecture; use edge computing nodes to perform preliminary cleaning, denoising and standardization on heterogeneous data to generate structured data streams; transmit the structured data streams to the central data processing unit, use deep learning algorithms to extract data features and understand semantics, and construct multidimensional data feature vectors; based on the multidimensional data feature vectors, use transfer learning methods to establish a cross-domain data association model to achieve semantic alignment and fusion of heterogeneous data;

[0012] Using graph database technology, we build a knowledge graph of the production process and map the fused multi-source heterogeneous data to the entities and relationships of the knowledge graph. Based on the time-series graph algorithm, we build dynamic time-series associations in the knowledge graph to form a traceable production process chain. Based on blockchain technology, we create tamper-proof digital fingerprints for key nodes and transactions in the production process and associate the digital fingerprints with the corresponding entities in the knowledge graph. We build an anomaly detection model based on graph neural networks to monitor the data flow and relationship changes in the knowledge graph in real time and identify potential production anomalies and quality risks.

[0013] Based on the knowledge graph and anomaly detection model, a transaction trigger mechanism is constructed to automatically identify production transactions that require collaborative processing and obtain transaction collaboration records. A multi-agent collaborative decision-making model is established using a reinforcement learning algorithm. This model can adaptively allocate tasks and resources based on the current production status and historical experience. Through the sequence-to-sequence model enhanced by the attention mechanism, unstructured collaboration requirements are converted into executable workflows. Based on the traceable production process chain and transaction collaboration records, multi-dimensional production analysis reports are automatically generated.

[0014] In an optional embodiment,

[0015] Based on the multi-dimensional data feature vector, a transfer learning method is used to establish a cross-domain data association model to achieve semantic alignment and fusion of heterogeneous data, including:

[0016] Use transfer learning methods to align feature vectors from different domains, including:

[0017] Select feature vectors of the source domain and the target domain; construct a domain adversarial neural network, including a feature extractor; train the domain adversarial neural network to minimize the difference in feature distribution between the source domain and the target domain; use the trained feature extractor to obtain the aligned feature vectors;

[0018] The cosine similarity is used to calculate the similarity between the aligned feature vectors, and the local sensitive hashing algorithm is used to accelerate the similarity calculation between large-scale vectors to obtain the similarity matrix;

[0019] Based on the similarity matrix, a spectral clustering algorithm is used to cluster the cross-domain data, including:

[0020] Construct an affinity matrix; calculate a Laplace matrix; calculate the eigenvector of the Laplace matrix; perform k-means clustering on the eigenvector, establish associations between data, and achieve heterogeneous data fusion.

[0021] In an optional embodiment,

[0022] Using graph database technology, we build a knowledge graph of the production process and map the fused multi-source heterogeneous data to the entities and relationships of the knowledge graph. Based on the time-series graph algorithm, we establish dynamic time-series associations in the knowledge graph to form a traceable production process chain, including:

[0023] Design a knowledge graph model and define the entity types and relationship types of the production process knowledge graph. The entity types include products, processes, equipment, raw materials, quality indicators, and operators. The relationship types include inclusion, use, production, monitoring, operation, and predecessor.

[0024] Select and configure the graph database platform, including installing the database server, creating a database instance, configuring memory and storage parameters, and enabling necessary plug-ins; then pre-process the fused multi-source heterogeneous data, including data cleaning, standardization, entity recognition, and relationship extraction;

[0025] Then, the preprocessed data is imported into the graph database using the graph database query language to create nodes and relationships and construct an initial knowledge graph. In the constructed initial knowledge graph, timestamp attributes are added to nodes and relationships, and a "happens after" temporal relationship is created based on the timestamp to establish a dynamic temporal association.

[0026] Based on the established temporal relationship, a temporal traversal algorithm is developed for temporal path query in the knowledge graph. By utilizing the constructed knowledge graph and temporal relationship, forward tracing, reverse tracing and quality problem tracing functions are realized to form a traceable production process chain.

[0027] In an optional embodiment,

[0028] Based on blockchain technology, tamper-proof digital fingerprints are created for key nodes and transactions in the production process, and the digital fingerprints are associated with the corresponding entities in the knowledge graph, including:

[0029] Design and deploy a permissioned blockchain network based on an open source blockchain distributed ledger, define the network topology, configure channels corresponding to production lines or product types, and write smart contracts for digital fingerprint creation, storage, and verification; use a multi-hash algorithm to generate digital fingerprints, including combining production process data, timestamps, and random numbers, first performing a first hash operation using the SHA256 algorithm, and then performing a second hash operation on the first hash result using the SHA3-256 algorithm to obtain the final digital fingerprint; implement the digital fingerprint storage function in the smart contract, encapsulate the entity ID, digital fingerprint, and timestamp into a structure, and store them in the blockchain world state after serialization;

[0030] Add digital fingerprint attributes to entities in the knowledge graph, create new nodes representing blockchain transactions, and establish the relationship between entity nodes and blockchain transaction nodes; develop a digital fingerprint verification process, first retrieve the digital fingerprint information of the target entity from the knowledge graph, then call the verification function of the blockchain smart contract, pass in the entity ID and digital fingerprint, and finally return the verification result; implement an integrated query interface for the knowledge graph and blockchain, and support simultaneous acquisition of entity information in the knowledge graph and digital fingerprint verification results in the blockchain based on the entity ID.

[0031] In an optional embodiment,

[0032] Using reinforcement learning algorithms, a multi-agent collaborative decision-making model is established. The model can adaptively allocate tasks and resources based on the current production status and historical experience. Through the sequence-to-sequence model enhanced by the attention mechanism, unstructured collaborative requirements are converted into executable workflows, including:

[0033] Construct a multi-agent deep deterministic policy gradient model based on the Actor-Critic architecture, where each agent represents a workstation or equipment on the production line. Each agent includes an actor network for generating actions and a critic network for evaluating the value of actions. Define a state space containing the current workstation task queue, resource utilization, and the status of adjacent workstations, as well as an action space containing task allocation decisions and resource scheduling strategies. Design a reward function that comprehensively considers production efficiency, resource utilization, and task completion quality.

[0034] Using the multi-agent deep deterministic policy gradient model, select the optimal action for each workstation according to the observed current state, and perform corresponding task allocation and resource scheduling operations; calculate the reward value after executing the action, and use the gained experience to update the multi-agent deep deterministic policy gradient model;

[0035] Construct an attention mechanism-enhanced sequence-to-sequence model including a bidirectional LSTM long short-term memory encoder, an attention layer, and a unidirectional LSTM decoder; convert the unstructured collaborative requirement text into a word sequence and encode it, input it into the attention mechanism-enhanced sequence-to-sequence model, and generate the corresponding workflow sequence; decode the generated workflow sequence into a readable workflow description;

[0036] According to the decoded workflow description, the updated multi-agent deep deterministic policy gradient model is used for task allocation and resource scheduling. During the execution of the production process, the latest production status information and historical experience data are continuously obtained, and the multi-agent deep deterministic policy gradient model is dynamically updated to achieve adaptive task allocation and resource scheduling optimization.

[0037] In an optional embodiment,

[0038] Based on the knowledge graph and anomaly detection model, a transaction trigger mechanism is built to automatically identify production transactions that need to be processed collaboratively. The transaction collaboration records obtained include:

[0039] Design a multi-layer autoencoder network structure as an anomaly detection model, use normal production data to train the autoencoder, calculate the reconstruction error and set a threshold for anomaly determination; collect sensor data and production status information from the production system in real time, and input the collected data into the trained anomaly detection model for analysis;

[0040] When an anomaly is detected, the knowledge graph is used for context analysis to determine the impact range and possible causes of the anomaly. Based on the anomaly and the results of the knowledge graph analysis, it is determined whether collaborative processing needs to be initiated. For transactions that require collaborative processing, a transaction collaborative record containing timestamps, anomaly information, affected equipment, affected processes, and collaborative requirements is generated.

[0041] The generated transaction collaboration records are input into the multi-agent collaborative decision-making system for task allocation and resource scheduling. The knowledge graph and anomaly detection model are continuously updated during the production process, and the anomaly detection threshold and collaborative processing strategy are dynamically adjusted. The knowledge graph is supplemented and corrected based on the collaborative processing results and production effects, and the anomaly detection model is incrementally learned using the newly added abnormal samples.

[0042] According to a second aspect of the embodiments of the present invention,

[0043] Provide production traceability and transaction collaboration system of MES system, including:

[0044] The first unit is used to deploy a distributed data collection network to collect heterogeneous data in real time from production equipment, environmental sensors, human-computer interaction terminals and enterprise information systems; classify and annotate the collected heterogeneous data to establish a unified data model architecture; use edge computing nodes to perform preliminary cleaning, denoising and standardization on heterogeneous data to generate structured data streams; transmit the structured data streams to the central data processing unit, use deep learning algorithms to extract data features and understand semantics, and construct multidimensional data feature vectors; based on the multidimensional data feature vectors, use transfer learning methods to establish a cross-domain data association model to achieve semantic alignment and fusion of heterogeneous data;

[0045] The second unit is used to use graph database technology to build a knowledge graph of the production process, map the fused multi-source heterogeneous data to the entities and relationships of the knowledge graph; based on the time series graph algorithm, establish dynamic time series associations in the knowledge graph to form a traceable production process chain; based on blockchain technology, create tamper-proof digital fingerprints for key nodes and transactions in the production process, and associate the digital fingerprints with the corresponding entities in the knowledge graph; build an anomaly detection model based on graph neural network to monitor the data flow and relationship changes in the knowledge graph in real time, and identify potential production anomalies and quality risks;

[0046] The third unit is used to build a transaction trigger mechanism based on the knowledge graph and anomaly detection model, automatically identify production transactions that need to be processed collaboratively, and obtain transaction collaboration records; use the reinforcement learning algorithm to establish a multi-agent collaborative decision-making model, which can adaptively allocate tasks and resources according to the current production status and historical experience; through the sequence-to-sequence model enhanced by the attention mechanism, convert unstructured collaboration requirements into executable workflows; based on the traceable production process chain and transaction collaboration records, automatically generate multi-dimensional production analysis reports.

[0047] According to a third aspect of the embodiments of the present invention,

[0048] An electronic device is provided, comprising:

[0049] processor;

[0050] a memory for storing processor-executable instructions;

[0051] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0052] According to a fourth aspect of the embodiments of the present invention,

[0053] 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.

[0054] Beneficial effects:

[0055] 1. Improve data integration and processing capabilities:

[0056] The present invention adopts a distributed data acquisition network and edge computing technology to realize the real-time acquisition and preliminary processing of heterogeneous data. Through deep learning algorithms and transfer learning methods, multidimensional data feature vectors and cross-domain data association models are constructed, which effectively solves the problem of data islands and realizes the semantic alignment and fusion of heterogeneous data. This method significantly improves the system's ability to process massive heterogeneous data, laying a solid data foundation for real-time monitoring and precise analysis of the production process.

[0057] 2. Enhance the integrity and credibility of production process traceability:

[0058] The present invention uses graph database technology to construct a knowledge graph of the production process, and combines it with a time-series graph algorithm to establish a dynamic time-series association, thus forming a complete traceable production process chain. At the same time, by introducing blockchain technology to create tamper-proof digital fingerprints for key nodes and transactions, the authenticity and integrity of the traceability data are greatly improved. This innovative traceability mechanism not only improves the accuracy of quality management, but also provides reliable technical support for responsibility identification, effectively reducing production risks.

[0059] 3. Realize intelligent collaborative decision-making:

[0060] The present invention constructs an intelligent transaction triggering mechanism based on knowledge graph and graph neural network, which can automatically identify production transactions that need to be processed collaboratively. The multi-agent collaborative decision-making model combined with the reinforcement learning algorithm realizes the adaptive allocation of tasks and resources. The sequence-to-sequence model enhanced by the attention mechanism converts unstructured collaborative requirements into executable workflows. This series of innovative technologies greatly enhances the system's intelligent collaborative decision-making capabilities, significantly improves production efficiency and emergency response capabilities, and provides strong support for enterprises to achieve flexible production and intelligent manufacturing. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 It is a flowchart of the production tracing and transaction coordination method of the MES system according to an embodiment of the present invention;

[0062] Figure 2 It is a structural diagram of the production traceability and transaction collaboration system of the MES system according to an embodiment of the present invention. DETAILED DESCRIPTION

[0063] 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.

[0064] 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.

[0065] Figure 1 FIG. 1 is a flow chart of a production traceability and transaction collaboration method of an MES system according to an embodiment of the present invention. Figure 1 As shown, the method includes:

[0066] S101. Deploy a distributed data collection network to collect heterogeneous data from production equipment, environmental sensors, human-computer interaction terminals and enterprise information systems in real time; classify and annotate the collected heterogeneous data and establish a unified data model architecture; use edge computing nodes to perform preliminary cleaning, denoising and standardization on heterogeneous data to generate structured data streams; transmit the structured data streams to the central data processing unit, use deep learning algorithms to extract data features and understand semantics, and construct multidimensional data feature vectors; based on the multidimensional data feature vectors, use transfer learning methods to establish a cross-domain data association model to achieve semantic alignment and fusion of heterogeneous data;

[0067] S102. Use graph database technology to build a knowledge graph of the production process, map the fused multi-source heterogeneous data to the entities and relationships of the knowledge graph; based on the time-series graph algorithm, establish dynamic time-series associations in the knowledge graph to form a traceable production process chain; based on blockchain technology, create tamper-proof digital fingerprints for key nodes and transactions in the production process, and associate the digital fingerprints with the corresponding entities in the knowledge graph; build an anomaly detection model based on graph neural networks to monitor the data flow and relationship changes in the knowledge graph in real time, and identify potential production anomalies and quality risks;

[0068] S103. Based on the knowledge graph and anomaly detection model, a transaction trigger mechanism is constructed to automatically identify production transactions that need to be collaboratively processed and obtain transaction collaboration records; a multi-agent collaborative decision-making model is established using a reinforcement learning algorithm, which can adaptively allocate tasks and resources based on the current production status and historical experience; through a sequence-to-sequence model enhanced by an attention mechanism, unstructured collaboration requirements are converted into executable workflows; based on a traceable production process chain and transaction collaboration records, a multi-dimensional production analysis report is automatically generated.

[0069] In this embodiment, the production traceability and transaction collaboration method of the MES (Manufacturing Execution System) system includes the following specific steps:

[0070] First, deploy a distributed data acquisition network. The network consists of multiple edge nodes, which are deployed at key locations on the production line. Each edge node is connected to multiple data sources, including PLC controllers of production equipment, environmental monitoring sensors, and human-computer interaction terminals used by operators. Raw data is collected in real time from these heterogeneous data sources through standardized industrial communication protocols (such as OPC UA, Modbus TCP, etc.).

[0071] Next, the collected heterogeneous data is classified and annotated. According to the type of data source, collection location, data format and other characteristics, the original data is divided into categories such as equipment data, environmental data, manual operation data, and management system data. A unified metadata model is formulated for each type of data, including attribute fields such as data description, collection time, source equipment, and unit. On this basis, a unified data model architecture for cross-domain data association is constructed.

[0072] Then, the edge computing nodes are used to perform preliminary processing on heterogeneous data. Filtering algorithms are used to remove noise and outliers in the data, such as using median filtering for temperature data and wavelet denoising for vibration data. According to the characteristics of different data sources, corresponding data cleaning rules are designed, such as removing invalid zero values ​​in device data and completing missing values ​​in environmental data. Finally, the processed data is converted into a unified JSON format to generate a structured data stream.

[0073] The preprocessed structured data stream is transmitted to the central data processing unit. This unit adopts a distributed architecture and contains multiple parallel computing nodes. First, a deep learning algorithm is used to extract data features, such as using a convolutional neural network to extract visual features from image data and using a long short-term memory network to extract time features from time series data. Then, natural language processing technology is used to perform semantic understanding of text data and extract key entities and relationships. Finally, a multidimensional vector representing the characteristics of different data sources is obtained.

[0074] Based on the obtained multi-dimensional feature vectors, a cross-domain data association model is established using a transfer learning method. Specifically, a basic model is first trained in the source domain (such as equipment data), and then the model is transferred to the target domain (such as quality data) to adapt to the new data distribution through fine-tuning. This method can effectively utilize the commonalities between different data sources to achieve semantic alignment and fusion of heterogeneous data. For example, equipment vibration data can be associated with product quality data to discover potential causal relationships.

[0075] Next, we use graph database technology to build a knowledge graph of the production process. First, we define the ontology model of the graph, including entity types such as equipment, materials, personnel, and processes, as well as relationship types such as processing, testing, and transportation. Then, we map the fused multi-source heterogeneous data to the entities and relationships in the graph. For example, we map the equipment operation data to the attributes of the equipment entity, and map the operation records to the relationship between the personnel entity and the equipment entity. In this way, we can form a comprehensive semantic expression of the production process.

[0076] Based on the constructed knowledge graph, a time-series graph algorithm is used to establish dynamic time-series associations. The specific approach is to introduce the time dimension into the graph and mark each entity and relationship with a timestamp. Then, the graph traversal algorithm is used to track the changes in entity states and the evolution of relationships at different time points. For example, the complete process of a batch of materials from raw material storage, processing, inspection to finished product delivery can be tracked to form a traceable production process chain.

[0077] In order to ensure the authenticity and immutability of data, blockchain technology is used to create digital fingerprints for key nodes and transactions in the production process. Specifically, for important production events (such as key quality inspections, process parameter adjustments, etc.), the relevant data is hashed to obtain a unique digital summary. The summary is then recorded in the blockchain network to form an immutable chain of evidence. At the same time, the transaction hash on the blockchain is associated with the corresponding entity in the knowledge graph to achieve data traceability.

[0078] On this basis, an anomaly detection model based on graph neural network is constructed. The model uses entities in the knowledge graph as nodes and relationships as edges, and performs feature learning through multi-layer graph convolution. Normal production data is used during model training to learn the normal association patterns between entities in the production process. In practical applications, the model monitors the data flow and relationship changes in the knowledge graph in real time, and can identify anomalies that deviate from the normal pattern. For example, potential production anomalies and quality risks such as abnormal fluctuations in equipment parameters and deviations from process paths can be detected.

[0079] Based on the constructed knowledge graph and anomaly detection model, a transaction trigger mechanism is designed to automatically identify production transactions that need to be collaboratively processed. Specifically, when the anomaly detection model finds a potential problem, the system will automatically analyze the impact range and severity of the anomaly and determine whether collaborative processing needs to be initiated. At the same time, the system will also monitor the preset key performance indicators (such as production progress, quality qualification rate, etc.), and will also trigger collaborative transactions when the indicators fluctuate abnormally. The system will record the collaborative needs that are automatically identified to form a transaction collaboration record.

[0080] In order to efficiently handle collaborative transactions, a multi-agent collaborative decision-making model is established using a reinforcement learning algorithm. The model abstracts different production units (such as process, quality, equipment and other departments) into agents, and adaptively allocates tasks and resources based on the current production status and historical experience. The reward function design of the model considers multiple objectives, including production efficiency, quality level, cost control, etc. Through continuous learning and optimization, the model can make reasonable collaborative decisions in a complex and changing production environment.

[0081] For the collaborative requirements identified by the system, a sequence-to-sequence model enhanced by the attention mechanism is used to convert unstructured collaborative requirements into executable workflows. The model first encodes the collaborative requirement text, and then gradually generates standardized workflow steps based on the learned context information. The attention mechanism is introduced during the decoding process, allowing the model to focus on key information in the input text. For example, for the collaborative requirement of "device A has abnormal fluctuations, and the process department needs to assist in analysis", the model can generate a detailed workflow including shutdown, inspection, analysis, adjustment and other steps.

[0082] Finally, based on the traceable production process chain and transaction collaboration records, a multi-dimensional production analysis report is automatically generated. The report includes production efficiency analysis, quality trend analysis, abnormal cause analysis, collaborative effect evaluation and other aspects. The system extracts relevant data from the knowledge graph and uses statistical analysis and data mining methods for in-depth analysis. For example, the association rule mining algorithm can be used to discover the key factors affecting product quality from historical data; through time series analysis, the probability of equipment failure can be predicted. The generated analysis report is presented in a visual way to support managers in making data-driven decisions.

[0083] Through the above steps, this method realizes the full process management from data collection, processing, analysis to collaborative decision-making, providing manufacturing enterprises with an efficient and intelligent production traceability and transaction collaboration solution.

[0084] In an optional implementation, based on the multidimensional data feature vector, a transfer learning method is used to establish a cross-domain data association model to achieve semantic alignment and fusion of heterogeneous data, including:

[0085] Use transfer learning methods to align feature vectors from different domains, including:

[0086] Select feature vectors of the source domain and the target domain; construct a domain adversarial neural network, including a feature extractor; train the domain adversarial neural network to minimize the difference in feature distribution between the source domain and the target domain; use the trained feature extractor to obtain the aligned feature vectors;

[0087] The cosine similarity is used to calculate the similarity between the aligned feature vectors, and the local sensitive hashing algorithm is used to accelerate the similarity calculation between large-scale vectors to obtain the similarity matrix;

[0088] Based on the similarity matrix, a spectral clustering algorithm is used to cluster the cross-domain data, including:

[0089] Construct an affinity matrix; calculate a Laplace matrix; calculate the eigenvector of the Laplace matrix; perform k-means clustering on the eigenvector, establish associations between data, and achieve heterogeneous data fusion.

[0090] In this implementation, based on multidimensional data feature vectors, a transfer learning method is used to establish a cross-domain data association model to achieve semantic alignment and fusion of heterogeneous data. The specific steps are as follows:

[0091] First, the feature vectors of different domains are aligned using the transfer learning method. The feature vectors of the source domain and the target domain are selected. For example, the source domain can be the word vector of text data, and the target domain can be the visual feature vector of image data. A domain adversarial neural network is constructed, which includes a feature extractor, a domain classifier, and a label predictor. The feature extractor consists of a multi-layer fully connected network, which is used to extract the common feature representation of the source domain and the target domain data. The domain classifier is used to determine which domain the feature comes from, and the label predictor is used to predict the category label of the data.

[0092] Then, the domain adversarial neural network is trained to minimize the difference in feature distribution between the source domain and the target domain while maintaining the discriminability of the features. During the training process, the goal of the feature extractor is to deceive the domain classifier so that it cannot distinguish the domain source of the features, so that the extracted features have domain invariance. For example, a gradient reversal layer can be used to implement adversarial training. After multiple rounds of iterative training, the trained feature extractor is used to obtain the aligned feature vectors of the source domain and target domain data.

[0093] Next, the cosine similarity is used to calculate the similarity between the aligned feature vectors. The value range of cosine similarity is between -1 and 1, 1 means the same direction, 0 means orthogonal, and -1 means the opposite direction. For two n-dimensional vectors a and b, the cosine similarity can be obtained by calculating the dot product of the two vectors divided by the product of the vector modulus. In order to accelerate the similarity calculation between large-scale vectors, the local sensitive hashing (LSH) algorithm is used. The core idea of ​​LSH is to map similar data to the same "bucket". Only the similarity of the vectors in the same bucket needs to be calculated, which greatly reduces the amount of calculation. In the specific implementation, the random projection method can be used to generate multiple hash functions to map the feature vectors to different buckets. Finally, an N×N similarity matrix is ​​obtained, where N is the number of data samples.

[0094] Based on the obtained similarity matrix, the spectral clustering algorithm is used to cluster the cross-domain data. First, the affinity matrix is ​​constructed. The similarity matrix can be used directly, or the Gaussian kernel function can be used to transform the similarity. Then the Laplace matrix is ​​calculated. There are many ways to choose, such as the undirected normalized Laplace matrix. Then the eigenvectors of the Laplace matrix are calculated, and the eigenvectors corresponding to the first k smallest eigenvalues ​​are selected to form an N×k matrix. Finally, k-means clustering is performed on each row of the matrix (representing a data point) to obtain the clustering results of the data.

[0095] Through the above steps, we can establish the association relationship between cross-domain data and realize the fusion of heterogeneous data. For example, we can cluster text data and image data into the same category to find the semantic association between them. This method can effectively process heterogeneous data of different modalities and structures, overcoming the limitation of traditional methods that it is difficult to directly compare heterogeneous data.

[0096] In practical applications, appropriate feature extraction methods and network structures can be selected according to specific tasks. For example, for text data, the pre-trained BERT (Bidirectional Encoder Representations from Transformers) model can be used as a feature extractor; for image data, the pre-trained CNN (convolution neural network) model can be used. In the training process of domain adversarial networks, the weights of each loss function need to be carefully adjusted to balance the domain invariance and discriminability of features. When using the LSH (Locality Sensitive Hashing) algorithm, the number of hash functions and the projection dimension are important hyperparameters that need to be weighed according to the data scale and computing resources. The number of clusters k in spectral clustering is also a key parameter. The optimal k value can be determined by observing the distribution of eigenvalues ​​or using indicators such as the silhouette coefficient.

[0097] This method can effectively mine potential associations between cross-domain data, providing important support for subsequent data analysis and decision-making. For example, in a recommendation system, users' text comments and product images can be associated to find the relationship between user preferences and product visual features, thereby providing more accurate personalized recommendations. In medical diagnosis, patients' medical records, medical images, and genetic data can be integrated to fully understand disease characteristics and assist doctors in making more accurate diagnoses.

[0098] In an optional implementation, graph database technology is used to construct a knowledge graph of the production process, and the fused multi-source heterogeneous data is mapped to the entities and relationships of the knowledge graph; based on the time sequence graph algorithm, dynamic time sequence associations are established in the knowledge graph to form a traceable production process chain including:

[0099] Design a knowledge graph model and define the entity types and relationship types of the production process knowledge graph. The entity types include products, processes, equipment, raw materials, quality indicators, and operators. The relationship types include inclusion, use, production, monitoring, operation, and predecessor.

[0100] Select and configure the graph database platform, including installing the database server, creating a database instance, configuring memory and storage parameters, and enabling necessary plug-ins; then pre-process the fused multi-source heterogeneous data, including data cleaning, standardization, entity recognition, and relationship extraction;

[0101] Then, the preprocessed data is imported into the graph database using the graph database query language to create nodes and relationships and construct an initial knowledge graph. In the constructed initial knowledge graph, timestamp attributes are added to nodes and relationships, and a "happens after" temporal relationship is created based on the timestamp to establish a dynamic temporal association.

[0102] Based on the established temporal relationship, a temporal traversal algorithm is developed for temporal path query in the knowledge graph. By utilizing the constructed knowledge graph and temporal relationship, forward tracing, reverse tracing and quality problem tracing functions are realized to form a traceable production process chain.

[0103] In order to realize the production process traceability method based on knowledge graph, it is necessary to design the knowledge graph model first. Define the entity types of the production process knowledge graph, including products, processes, equipment, raw materials, quality indicators and operators. At the same time, define the relationship types, including containment, use, production, monitoring, operation and predecessor. For example, you can define the "product" entity type to contain attributes such as product ID, name, specification, etc.; the "process" entity type to contain process ID, name, description and other attributes; the "equipment" entity type to contain equipment ID, name, model and other attributes. Relationship types such as "containment" are used to indicate the processes included in the product, "use" indicates the equipment and raw materials used in the process, and "production" indicates the products produced by the process.

[0104] Next, select and configure the graph database platform. You can choose Neo4j as the graph database platform, install the Neo4j server, create a database instance named "ProductionKG", configure the memory to 8GB, and the storage space to 100GB. Enable the APOC plug-in to support advanced graph algorithms. Preprocess the fused multi-source heterogeneous data, including removing duplicates and outliers, unifying the date format, identifying entities such as products and processes, and extracting the relationships between entities. For example, identify product entities from order data extracted from the ERP system, identify process and equipment entities from production records extracted from the MES system, and extract the before and after relationships between processes.

[0105] Use Cypher query language to import preprocessed data into Neo4j database. Build initial knowledge graph containing various entities and relationships through batch import.

[0106] In the initial knowledge graph, add timestamp attributes to nodes and relationships. For example, add start time and end time attributes to the process node:

[0107] MATCH (s:Step {id:'S001'});

[0108] SET s.startTime = '2023-05-01 09:00:00', s.endTime = '2023-05-01 11:00:00';

[0109] Create a "happens after" timing relationship based on timestamps and establish dynamic timing associations:

[0110] MATCH (s1:Step), (s2:Step);

[0111] WHERE s1.endTime < s2.startTime;

[0112] CREATE (s1)-[:HAPPENED_BEFORE]->(s2);

[0113] Based on the established temporal relationship, a temporal traversal algorithm is developed for temporal path query. A depth-first search algorithm can be used, starting from the starting node, traversing backward along the "HAPPENED_BEFORE" relationship, and recording the traversal path until the termination condition is reached.

[0114] By using the knowledge graph and time-series relationship, forward tracing, reverse tracing and quality problem tracing functions are realized. Forward tracing starts from raw materials, traverses forward along the "use" and "production" relationships, and tracks the use of raw materials to the final product. Reverse tracing starts from the final product, traverses backward along the "include" and "use" relationships, and traces the production process of the product to the raw materials. Quality problem tracing starts from the problem product, combines forward and reverse tracing, locates possible problem sources and analyzes the scope of impact.

[0115] The query returns the complete production path from raw materials to the problematic product, forming a traceable production process chain. In this way, the source of the problem can be quickly located and the scope of impact can be assessed, effectively supporting quality management and problem handling.

[0116] In an optional implementation, based on blockchain technology, creating tamper-proof digital fingerprints for key nodes and transactions in the production process, and associating the digital fingerprints with corresponding entities in the knowledge graph includes:

[0117] Design and deploy a permissioned blockchain network based on an open source blockchain distributed ledger, define the network topology, configure channels corresponding to production lines or product types, and write smart contracts for digital fingerprint creation, storage, and verification; use a multi-hash algorithm to generate digital fingerprints, including combining production process data, timestamps, and random numbers, first performing a first hash operation using the SHA256 algorithm, and then performing a second hash operation on the first hash result using the SHA3-256 algorithm to obtain the final digital fingerprint; implement the digital fingerprint storage function in the smart contract, encapsulate the entity ID, digital fingerprint, and timestamp into a structure, and store them in the blockchain world state after serialization;

[0118] Add digital fingerprint attributes to entities in the knowledge graph, create new nodes representing blockchain transactions, and establish the relationship between entity nodes and blockchain transaction nodes; develop a digital fingerprint verification process, first retrieve the digital fingerprint information of the target entity from the knowledge graph, then call the verification function of the blockchain smart contract, pass in the entity ID and digital fingerprint, and finally return the verification result; implement an integrated query interface for the knowledge graph and blockchain, and support simultaneous acquisition of entity information in the knowledge graph and digital fingerprint verification results in the blockchain based on the entity ID.

[0119] The specific implementation method of creating tamper-proof digital fingerprints for key nodes and transactions in the production process based on blockchain technology and associating the digital fingerprints with the corresponding entities in the knowledge graph is as follows:

[0120] First, design and deploy a permissioned blockchain network based on an open source blockchain distributed ledger. Define the network topology, including multiple organizations (such as manufacturers, suppliers, regulators, etc.), multiple nodes (such as production line nodes, warehouse nodes, etc.), and corresponding channels. Taking the smart manufacturing production line as an example, independent channels can be configured for different product types (such as mobile phones, TVs, etc.) to achieve data isolation. For example, the mobile phone production channel can include nodes of related organizations such as mobile phone manufacturers and major component suppliers.

[0121] Next, write a smart contract to implement the creation, storage and verification of digital fingerprints. Smart contracts are developed in Go language and mainly include the following functions: createFingerprint is used to create digital fingerprints, storeFingerprint is used to store digital fingerprints, and verifyFingerprint is used to verify digital fingerprints. In the createFingerprint function, first obtain the production process data (such as product ID, batch number, production parameters, etc.), the current timestamp, and a random number. Then concatenate these data into a string and use the SHA256 algorithm to calculate the first hash value. Then use the SHA3-256 algorithm to perform a second hash operation on the first hash value to finally obtain the digital fingerprint.

[0122] Take mobile phone production as an example. Assume the following production data: product ID is "PHONE001", batch number is "BATCH20230501", production parameters are "{"temperature": 25, "pressure": 1.0}", timestamp is "1682899200", random number is "123456". After concatenating these data, we get the string:

[0123] "PHONE001BATCH20230501

[0124] {"temperature": 25, "pressure": 1.0}1682899200123456".

[0125] Perform SHA256 hashing on the string to get the first hash value:

[0126] "8f4e8d9a1b3c2d5e6f7a8b9c0d1e2f3a4b5c6d7e8f9a0b1c2d3e4f5a6b7c8d9".

[0127] Then perform SHA3-256 hashing on the first hash value to finally get the digital fingerprint:

[0128] "2a1b3c4d5e6f7a8b9c0d1e2f3a4b5c6d7e8f9a0b1c2d3e4f5a6b7c8d9e0f1".

[0129] In the storeFingerprint function, the entity ID, digital fingerprint and timestamp are encapsulated into a structure, and then serialized and stored in the blockchain world state.

[0130] In the knowledge graph, digital fingerprint attributes are added to the entity, and a new node representing the blockchain transaction is created. In this way, digital fingerprint attributes are added to the product entity in the knowledge graph, a new node representing the blockchain transaction is created, and the relationship between the product entity and the blockchain transaction node is established.

[0131] Next, we develop the digital fingerprint verification process. First, we retrieve the digital fingerprint information of the target entity from the knowledge graph, which can be achieved through the following Cypher query statement:

[0132] MATCH (p:Product {id: 'PHONE001'});

[0133] RETURN p.digitalFingerprint;

[0134] After obtaining the digital fingerprint, call the verifyFingerprint function of the blockchain smart contract for verification. In the verifyFingerprint function, first obtain the stored digital fingerprint data from the world state according to the entity ID, and then compare it with the incoming digital fingerprint. If it matches, the result of successful verification is returned; otherwise, the result of failed verification is returned.

[0135] Finally, the integrated query interface of knowledge graph and blockchain is implemented. A RESTful API is developed to support the simultaneous acquisition of entity information in the knowledge graph and digital fingerprint verification results in the blockchain based on entity ID. The processing flow of the API is as follows:

[0136] 1. Receive an API request containing an entity ID.

[0137] 2. Query the knowledge graph to obtain the attribute information of the entity.

[0138] 3. Obtain the digital fingerprint of the entity from the knowledge graph.

[0139] 4. Call the verifyFingerprint function of the blockchain smart contract to verify the digital fingerprint.

[0140] 5. Combine the knowledge graph query results and blockchain verification results and return them to the client.

[0141] Through the above steps, we have realized the complete process of creating tamper-proof digital fingerprints for key nodes and transactions in the production process based on blockchain technology, and associating the digital fingerprints with the corresponding entities in the knowledge graph. This method combines the tamper-proof characteristics of blockchain and the semantic association ability of knowledge graph, providing a strong guarantee for the authenticity and traceability of data in the intelligent manufacturing process.

[0142] In an optional implementation, a multi-agent collaborative decision-making model is established using a reinforcement learning algorithm, which can adaptively allocate tasks and resources based on the current production status and historical experience; a sequence-to-sequence model enhanced by an attention mechanism is used to convert unstructured collaborative requirements into executable workflows, including:

[0143] Construct a multi-agent deep deterministic policy gradient model based on the Actor-Critic architecture, where each agent represents a workstation or equipment on the production line. Each agent includes an actor network for generating actions and a critic network for evaluating the value of actions. Define a state space containing the current workstation task queue, resource utilization, and the status of adjacent workstations, as well as an action space containing task allocation decisions and resource scheduling strategies. Design a reward function that comprehensively considers production efficiency, resource utilization, and task completion quality.

[0144] Using the multi-agent deep deterministic policy gradient model, select the optimal action for each workstation according to the observed current state, and perform corresponding task allocation and resource scheduling operations; calculate the reward value after executing the action, and use the gained experience to update the multi-agent deep deterministic policy gradient model;

[0145] Construct an attention-enhanced sequence-to-sequence model including a bidirectional LSTM (Long Short-Term Memory) encoder, an attention layer, and a unidirectional LSTM decoder; convert the unstructured collaborative requirement text into a word sequence and encode it, input it into the attention-enhanced sequence-to-sequence model, and generate the corresponding workflow sequence; decode the generated workflow sequence into a readable workflow description;

[0146] According to the decoded workflow description, the updated multi-agent deep deterministic policy gradient model is used for task allocation and resource scheduling. During the execution of the production process, the latest production status information and historical experience data are continuously obtained, and the multi-agent deep deterministic policy gradient model is dynamically updated to achieve adaptive task allocation and resource scheduling optimization.

[0147] According to the method, a multi-agent deep deterministic policy gradient model based on the Actor-Critic architecture is first constructed. Each agent in the model represents a workstation or device on the production line, including an actor network for generating actions and a critic network for evaluating the value of actions. In specific implementation, a deep neural network can be used as a function approximator for the actor and the critic. For example, the actor network can use a 3-layer fully connected network, the number of neurons in the input layer is the same as the dimension of the state space, the hidden layer uses 64 neurons, and the number of neurons in the output layer is the same as the dimension of the action space; the critic network also uses a 3-layer fully connected network, the input layer includes states and actions, the hidden layer uses 64 neurons, and the output layer is 1 neuron to represent the state-action value function.

[0148] Next, define the state space and action space. The state space contains the current workstation task queue, resource utilization, and neighboring workstation status. For example, the task queue can be represented by a vector of length 10, with each element representing the number of tasks of a type; resource utilization can be represented by a scalar between 0 and 1; and the neighboring workstation status can be represented by a vector of the same dimension as the current workstation. The action space contains task allocation decisions and resource scheduling strategies, and can be designed as a discrete action space, for example, a combination of 10 predefined task allocation schemes and 5 resource scheduling strategies.

[0149] Then design a reward function that comprehensively considers production efficiency, resource utilization, and task completion quality. The reward function can be defined as the weighted sum of these three indicators, and the weights can be adjusted according to actual production needs. For example, production efficiency can be expressed by the number of tasks completed per unit time, resource utilization can be expressed by the average utilization of various resources, and task completion quality can be expressed by the pass rate.

[0150] Using the constructed multi-agent deep deterministic policy gradient model, the optimal action is selected for each workstation based on the observed current state. In the specific operation, the state information of the current workstation is first obtained and input into the actor network of the corresponding agent to obtain a deterministic action output. Then the action is executed and the corresponding task allocation and resource scheduling operations are performed. For example, if the selected action is to assign 3 Class A tasks to workstation No. 1 and increase the resource allocation by 50%, the task queue and resource allocation of workstation No. 1 are updated accordingly.

[0151] After executing the action, calculate the reward value obtained. According to the previously defined reward function, calculate the production efficiency, resource utilization and task completion quality after executing the action, and get the comprehensive reward value according to the set weights. Store the information such as the state before executing the action, the executed action, the reward obtained, and the new state after executing the action as an experience data. Use the collected experience data to update the multi-agent deep deterministic policy gradient model. The specific update process includes updating the critic network to reduce the TD error, and then using the policy gradient method to update the actor network to maximize the expected return.

[0152] Next, we build a sequence-to-sequence model enhanced by the attention mechanism to transform unstructured collaborative requirements into executable workflows. The model contains a bidirectional LSTM encoder, an attention layer, and a unidirectional LSTM decoder. The encoder uses a bidirectional LSTM to capture the contextual information of the input sequence; the decoder uses a unidirectional LSTM to predict the next output based on the generated output sequence. The attention layer is used to dynamically focus on different parts of the input sequence during the decoding process.

[0153] Convert unstructured collaborative demand text into word sequences and encode them. For example, convert a demand text such as "need to complete the production of 100 Class A products and 50 Class B products within 24 hours" into a word sequence such as ["need", "in", "24", "hours", "within", "complete", "100", "pieces", "Class A", "products", "and", "50", "pieces", "Class B", "products", "of", "production"]. Then use a pre-trained word vector model (such as Word2Vec) to convert the word sequence into a vector sequence as the input of the encoder.

[0154] The encoded vector sequence is input into the sequence-to-sequence model enhanced by the attention mechanism to generate the corresponding workflow sequence. During the decoding process, the model predicts the next output word based on the partial output generated and the context vector calculated by the attention mechanism. The generated workflow sequence may be in the form of ["allocation", "50", "pieces", "Class A", "task", "give", "workstation one", "allocation", "25", "pieces", "Class B", "task", "give", "workstation two", ...].

[0155] Decode the generated workflow sequence into a readable workflow description. This step can be achieved by a simple template filling method, filling the key information in the generated sequence into the predefined workflow description template. For example, you can get a workflow description like "First assign 50 class A tasks to workstation number one, then assign 25 class B tasks to workstation number two,..."

[0156] According to the decoded workflow description, the updated multi-agent deep deterministic policy gradient model is used for task allocation and resource scheduling. In specific operations, each task allocation step in the workflow is converted into a decision cycle of the model. In each decision cycle, the optimal action is selected according to the current state, the corresponding task allocation and resource scheduling operations are performed, and the system state is updated.

[0157] During the production process, the latest production status information and historical experience data are continuously obtained to dynamically update the multi-agent deep deterministic policy gradient model. A fixed update cycle can be set, for example, a model update is performed after every 100 decision cycles. During the update, the new experience data collected is used to update the critic network and the actor network according to the method described above. In this way, the model can continuously adapt to the changing production environment and achieve adaptive task allocation and resource scheduling optimization.

[0158] In an optional implementation, based on the knowledge graph and the anomaly detection model, a transaction trigger mechanism is constructed to automatically identify production transactions that need to be collaboratively processed, and the transaction collaboration records obtained include:

[0159] Design a multi-layer autoencoder network structure as an anomaly detection model, use normal production data to train the autoencoder, calculate the reconstruction error and set a threshold for anomaly determination; collect sensor data and production status information from the production system in real time, and input the collected data into the trained anomaly detection model for analysis;

[0160] When an anomaly is detected, the knowledge graph is used for context analysis to determine the impact range and possible causes of the anomaly. Based on the anomaly and the results of the knowledge graph analysis, it is determined whether collaborative processing needs to be initiated. For transactions that require collaborative processing, a transaction collaborative record containing timestamps, anomaly information, affected equipment, affected processes, and collaborative requirements is generated.

[0161] The generated transaction collaboration records are input into the multi-agent collaborative decision-making system for task allocation and resource scheduling. The knowledge graph and anomaly detection model are continuously updated during the production process, and the anomaly detection threshold and collaborative processing strategy are dynamically adjusted. The knowledge graph is supplemented and corrected based on the collaborative processing results and production effects, and the anomaly detection model is incrementally learned using the newly added abnormal samples.

[0162] In order to build a transaction trigger mechanism based on knowledge graph and anomaly detection model to automatically identify production transactions that need to be processed collaboratively, the following specific implementation methods can be adopted:

[0163] First, a multi-layer autoencoder network structure is designed as an anomaly detection model. The autoencoder contains an input layer, multiple hidden layers, and an output layer. The number of neurons in the input layer is the same as the total number of sensors and state variables in the production system, for example, it can be set to 100 neurons. The hidden layer adopts a 3-layer structure, with 80, 50, and 80 neurons respectively. The number of neurons in the output layer is the same as that in the input layer. ReLU is selected as the activation function, and the mean square error is used as the loss function.

[0164] Next, use normal production data to train the autoencoder. Filter out normal operating data within a period of time from the historical data of the production system, for example, select normal production data from the last three months. Divide these data into a training set and a validation set at a ratio of 80% and 20%. Use the stochastic gradient descent algorithm to train the autoencoder, set the learning rate to 0.001, the batch size to 64, and the number of training rounds to 100. Calculate the reconstruction error on the validation set, and take the 99.9% quantile as the threshold for abnormal judgment, for example, set the threshold to 0.05.

[0165] Then, sensor data and production status information are collected from the production system in real time. Data can be collected every 1 minute, including sensor values ​​such as temperature, pressure, flow, as well as information such as equipment operation status and production parameters. The collected data is standardized and input into the trained anomaly detection model to calculate the reconstruction error. If the reconstruction error exceeds the preset threshold of 0.05, it is judged as an abnormal situation.

[0166] When an anomaly is detected, the system uses a pre-built knowledge graph to perform contextual analysis to determine the impact and possible causes of the anomaly. The knowledge graph contains entities such as equipment, processes, parameters, faults, and the relationships between them. For example, when a temperature sensor value is abnormal, the knowledge graph can quickly locate the device where the sensor is located, the related processes, and the type of fault that may cause the temperature anomaly. The system tracks the propagation path of the anomaly and evaluates the potential impact on downstream processes and product quality.

[0167] Based on the abnormal situation and the results of knowledge graph analysis, the system will determine whether collaborative processing needs to be initiated. The judgment criteria include: the severity of the abnormality, the scope of impact, whether collaboration between multiple departments is required, etc. For transactions that require collaborative processing, the system will automatically generate a transaction collaboration record, which contains the following information: occurrence timestamp, abnormal sensor or device information, abnormal value, impact scope, possible causes, affected processes, and recommended collaborative processing solutions.

[0168] For example, once temperature sensor A detected an abnormal temperature rise to 95°C (normal range is 60-80°C). The transaction collaboration record generated by the system can be:

[0169] {

[0170] "Timestamp": "2023-06-01 10:30:15",

[0171] "Abnormal Device": "Temperature Sensor A",

[0172] "Outliers": "95°C",

[0173] "Affected Scope": ["Reactor No. 1", "Cooling System No. 1"],

[0174] "Possible causes": ["Cooling system failure", "Reaction out of control"],

[0175] "Affected Processes": ["Polymerization", "Product Cooling"],

[0176] "Collaborative requirements": ["Process engineer", "Equipment maintenance personnel", "Quality control personnel"]

[0177] }.

[0178] The generated transaction collaboration records will be input into the multi-agent collaborative decision-making system, which is responsible for task allocation and resource scheduling. The collaborative decision-making system will assign tasks to corresponding personnel or intelligent agents based on factors such as the urgency of the transaction and the required professional skills. For example, for the above-mentioned abnormal temperature event, the system may simultaneously notify the process engineer to evaluate the reaction, dispatch maintenance personnel to check the cooling system, and arrange quality control personnel to pay attention to changes in product quality.

[0179] During the production process, the system will continuously update the knowledge graph and anomaly detection model. After each collaborative processing is completed, the relevant experience and solutions will be added to the knowledge graph to enrich the knowledge base of anomaly handling. At the same time, the system will regularly (such as weekly) use the newly added production data to incrementally learn the anomaly detection model to adapt to changes in the production environment. In addition, the system will dynamically adjust the anomaly detection threshold according to the actual production situation. For example, if it is found that some false alarm rates are high, the judgment threshold of this type of anomaly can be appropriately increased.

[0180] Finally, the system will supplement and correct the knowledge graph based on the results of collaborative processing and production effects. For example, if a collaborative processing successfully solves a problem, the relevant processing steps and experience will be added to the knowledge graph. If it is found that some preset fault causes do not match the actual situation, the relationship in the knowledge graph will be corrected. At the same time, the system will use the newly added abnormal samples for incremental learning of the anomaly detection model to continuously improve the accuracy and generalization ability of the model.

[0181] Through the above steps, this method can effectively build a transaction trigger mechanism, realize timely discovery and intelligent collaborative processing of production anomalies, and thus improve the stability and efficiency of the production system.

[0182] Figure 2 FIG. 1 is a schematic diagram of the structure of the production traceability and transaction collaboration system of the MES system according to an embodiment of the present invention. Figure 2 As shown, the system comprises:

[0183] The first unit is used to deploy a distributed data collection network to collect heterogeneous data in real time from production equipment, environmental sensors, human-computer interaction terminals and enterprise information systems; classify and annotate the collected heterogeneous data to establish a unified data model architecture; use edge computing nodes to perform preliminary cleaning, denoising and standardization on heterogeneous data to generate structured data streams; transmit the structured data streams to the central data processing unit, use deep learning algorithms to extract data features and understand semantics, and construct multidimensional data feature vectors; based on the multidimensional data feature vectors, use transfer learning methods to establish a cross-domain data association model to achieve semantic alignment and fusion of heterogeneous data;

[0184] The second unit is used to use graph database technology to build a knowledge graph of the production process, map the fused multi-source heterogeneous data to the entities and relationships of the knowledge graph; based on the time series graph algorithm, establish dynamic time series associations in the knowledge graph to form a traceable production process chain; based on blockchain technology, create tamper-proof digital fingerprints for key nodes and transactions in the production process, and associate the digital fingerprints with the corresponding entities in the knowledge graph; build an anomaly detection model based on graph neural network to monitor the data flow and relationship changes in the knowledge graph in real time, and identify potential production anomalies and quality risks;

[0185] The third unit is used to build a transaction trigger mechanism based on the knowledge graph and anomaly detection model, automatically identify production transactions that need to be processed collaboratively, and obtain transaction collaboration records; use the reinforcement learning algorithm to establish a multi-agent collaborative decision-making model, which can adaptively allocate tasks and resources according to the current production status and historical experience; through the sequence-to-sequence model enhanced by the attention mechanism, convert unstructured collaboration requirements into executable workflows; based on the traceable production process chain and transaction collaboration records, automatically generate multi-dimensional production analysis reports.

[0186] According to a third aspect of the embodiments of the present invention,

[0187] An electronic device is provided, comprising:

[0188] processor;

[0189] a memory for storing processor-executable instructions;

[0190] The processor is configured to call the instructions stored in the memory to execute the aforementioned method.

[0191] According to a fourth aspect of the embodiments of the present invention,

[0192] 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.

[0193] 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.

[0194] 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 with 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. The production traceability and transaction collaboration method of the MES system is characterized by: include: Deploy a distributed data collection network to collect heterogeneous data in real time from production equipment, environmental sensors, human-computer interaction terminals, and enterprise information systems; Classify and annotate the collected heterogeneous data and establish a unified data model architecture; Use edge computing nodes to perform preliminary cleaning, denoising, and standardization on heterogeneous data to generate structured data streams; The structured data stream is transmitted to the central data processing unit, and the data feature extraction and semantic understanding are performed using the deep learning algorithm to construct a multi-dimensional data feature vector; based on the multi-dimensional data feature vector, a transfer learning method is used to establish a cross-domain data association model to achieve semantic alignment and fusion of heterogeneous data; the semantic alignment and fusion of heterogeneous data is achieved based on domain adversarial neural network and spectral clustering algorithm; Using graph database technology, we build a knowledge graph of the production process and map the fused multi-source heterogeneous data to the entities and relationships of the knowledge graph. Based on the time-series graph algorithm, dynamic time-series associations are established in the knowledge graph to form a traceable production process chain; Based on blockchain technology, we create tamper-proof digital fingerprints for key nodes and transactions in the production process, and associate the digital fingerprints with the corresponding entities in the knowledge graph. We build an anomaly detection model based on graph neural networks to monitor the data flow and relationship changes in the knowledge graph in real time, and identify potential production anomalies and quality risks. Based on the knowledge graph and anomaly detection model, a transaction trigger mechanism is built to automatically identify production transactions that need to be processed collaboratively and obtain transaction collaboration records; Using reinforcement learning algorithms, a multi-agent collaborative decision-making model is established, which can adaptively allocate tasks and resources based on current production status and historical experience; Through the sequence-to-sequence model enhanced by the attention mechanism, unstructured collaboration requirements are transformed into executable workflows; based on the traceable production process chain and transaction collaboration records, multi-dimensional production analysis reports are automatically generated; the anomaly detection model is a multi-layer autoencoder network.

2. The method according to claim 1, characterized in that Based on the multi-dimensional data feature vector, a transfer learning method is used to establish a cross-domain data association model to achieve semantic alignment and fusion of heterogeneous data, including: Use transfer learning methods to align feature vectors from different domains, including: Select feature vectors of the source domain and the target domain; construct a domain adversarial neural network, including a feature extractor; train the domain adversarial neural network to minimize the difference in feature distribution between the source domain and the target domain; use the trained feature extractor to obtain the aligned feature vectors; The cosine similarity is used to calculate the similarity between the aligned feature vectors, and the local sensitive hashing algorithm is used to accelerate the similarity calculation between large-scale vectors to obtain the similarity matrix; Based on the similarity matrix, a spectral clustering algorithm is used to cluster the cross-domain data, including: Construct an affinity matrix; calculate a Laplace matrix; calculate the eigenvector of the Laplace matrix; perform k-means clustering on the eigenvector, establish associations between data, and achieve heterogeneous data fusion.

3. The method according to claim 1, characterized in that Using graph database technology, we build a knowledge graph of the production process and map the fused multi-source heterogeneous data to the entities and relationships of the knowledge graph. Based on the time series graph algorithm, dynamic time series associations are established in the knowledge graph to form a traceable production process chain including: Design a knowledge graph model and define the entity types and relationship types of the production process knowledge graph. The entity types include products, processes, equipment, raw materials, quality indicators, and operators. The relationship types include inclusion, use, production, monitoring, operation, and predecessor. Select and configure the graph database platform, including installing the database server, creating a database instance, configuring memory and storage parameters, and enabling plug-ins; then pre-process the fused multi-source heterogeneous data, including data cleaning, standardization, entity recognition, and relationship extraction; Then, the preprocessed data is imported into the graph database using the graph database query language to create nodes and relationships and construct an initial knowledge graph. In the constructed initial knowledge graph, timestamp attributes are added to nodes and relationships, and a "happens after" temporal relationship is created based on the timestamp to establish a dynamic temporal association. Based on the established temporal relationship, a temporal traversal algorithm is developed for temporal path query in the knowledge graph. By utilizing the constructed knowledge graph and temporal relationship, forward tracing, reverse tracing and quality problem tracing functions are realized to form a traceable production process chain.

4. The method according to claim 1, characterized in that Based on blockchain technology, tamper-proof digital fingerprints are created for key nodes and transactions in the production process, and the digital fingerprints are associated with the corresponding entities in the knowledge graph, including: Design and deploy a permissioned blockchain network based on Hyperledger Fabric, define the network topology, configure channels corresponding to production lines or product types, and write smart contracts for digital fingerprint creation, storage, and verification; use multiple hash algorithms to generate digital fingerprints, including combining production process data, timestamps, and random numbers, first performing the first hash operation using the SHA256 algorithm, and then performing the second hash operation on the first hash result using the SHA3-256 algorithm to obtain the final digital fingerprint; implement the digital fingerprint storage function in the smart contract, encapsulate the entity ID, digital fingerprint, and timestamp into a structure, and store it in the blockchain world state after serialization; Add digital fingerprint attributes to entities in the knowledge graph, create new nodes representing blockchain transactions, and establish the relationship between entity nodes and blockchain transaction nodes; develop a digital fingerprint verification process, first retrieve the digital fingerprint information of the target entity from the knowledge graph, then call the verification function of the blockchain smart contract, pass in the entity ID and digital fingerprint, and finally return the verification result; implement an integrated query interface for the knowledge graph and blockchain, and support simultaneous acquisition of entity information in the knowledge graph and digital fingerprint verification results in the blockchain based on the entity ID.

5. The method according to claim 1, characterized in that Using reinforcement learning algorithms, a multi-agent collaborative decision-making model is established, which can adaptively allocate tasks and resources based on current production status and historical experience; The sequence-to-sequence model enhanced by the attention mechanism transforms unstructured collaborative requirements into executable workflows including: Construct a multi-agent deep deterministic policy gradient model based on the Actor-Critic architecture, where each agent represents a workstation or equipment on the production line. Each agent includes an actor network for generating actions and a critic network for evaluating the value of actions. Define a state space containing the current workstation task queue, resource utilization, and the status of adjacent workstations, as well as an action space containing task allocation decisions and resource scheduling strategies. Design a reward function that comprehensively considers production efficiency, resource utilization, and task completion quality. Using the multi-agent deep deterministic policy gradient model, select the optimal action for each workstation according to the observed current state, and perform corresponding task allocation and resource scheduling operations; calculate the reward value after executing the action, and use the gained experience to update the multi-agent deep deterministic policy gradient model; Construct an attention-enhanced sequence-to-sequence model including a bidirectional LSTM encoder, an attention layer, and a unidirectional LSTM decoder; convert the unstructured collaborative requirement text into a word sequence and encode it, input it into the attention-enhanced sequence-to-sequence model, and generate the corresponding workflow sequence; decode the generated workflow sequence into a readable workflow description; According to the decoded workflow description, the updated multi-agent deep deterministic policy gradient model is used for task allocation and resource scheduling. During the execution of the production process, the latest production status information and historical experience data are continuously obtained, and the multi-agent deep deterministic policy gradient model is dynamically updated to achieve adaptive task allocation and resource scheduling optimization.

6. The method according to claim 1, characterized in that Based on the knowledge graph and anomaly detection model, a transaction trigger mechanism is built to automatically identify production transactions that need to be processed collaboratively. The transaction collaboration records obtained include: Design a multi-layer autoencoder network structure as an anomaly detection model, use normal production data to train the autoencoder, calculate the reconstruction error and set a threshold for anomaly determination; collect sensor data and production status information from the production system in real time, and input the collected data into the trained anomaly detection model for analysis; When an anomaly is detected, the knowledge graph is used for context analysis to determine the impact range and possible causes of the anomaly. Based on the anomaly and the results of the knowledge graph analysis, it is determined whether collaborative processing needs to be initiated. For transactions that require collaborative processing, a transaction collaborative record containing timestamps, anomaly information, affected equipment, affected processes, and collaborative requirements is generated. The generated transaction collaboration records are input into the multi-agent collaborative decision-making system for task allocation and resource scheduling. The knowledge graph and anomaly detection model are continuously updated during the production process, and the anomaly detection threshold and collaborative processing strategy are dynamically adjusted. The knowledge graph is supplemented and corrected based on the collaborative processing results and production effects, and the anomaly detection model is incrementally learned using the newly added abnormal samples.

7. A production traceability and transaction collaboration system of an MES system, used to implement the method of any one of claims 1 to 6, characterized in that: include: The first unit is used to deploy a distributed data collection network to collect heterogeneous data in real time from production equipment, environmental sensors, human-computer interaction terminals, and enterprise information systems; Classify and annotate the collected heterogeneous data and establish a unified data model architecture; Use edge computing nodes to perform preliminary cleaning, denoising, and standardization on heterogeneous data to generate structured data streams; Transmit the structured data stream to the central data processing unit, use the deep learning algorithm to extract data features and semantic understanding, and construct a multi-dimensional data feature vector; based on the multi-dimensional data feature vector, use the transfer learning method to establish a cross-domain data association model to achieve semantic alignment and fusion of heterogeneous data; The second unit is used to build a knowledge graph of the production process using graph database technology, and map the fused multi-source heterogeneous data to the entities and relationships of the knowledge graph; Based on the time-series graph algorithm, dynamic time-series associations are established in the knowledge graph to form a traceable production process chain; Based on blockchain technology, we create tamper-proof digital fingerprints for key nodes and transactions in the production process, and associate the digital fingerprints with the corresponding entities in the knowledge graph. We build an anomaly detection model based on graph neural networks to monitor the data flow and relationship changes in the knowledge graph in real time, and identify potential production anomalies and quality risks. The third unit is used to build a transaction trigger mechanism based on the knowledge graph and anomaly detection model, automatically identify production transactions that need to be collaboratively processed, and obtain transaction collaboration records; Using reinforcement learning algorithms, a multi-agent collaborative decision-making model is established, which can adaptively allocate tasks and resources based on current production status and historical experience; Through the sequence-to-sequence model enhanced by the attention mechanism, unstructured collaboration requirements are transformed into executable workflows; based on the traceable production process chain and transaction collaboration records, multi-dimensional production analysis reports are automatically generated.

8. 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 according to any one of claims 1 to 6.

9. 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 6 is implemented.

Images