Quality spectrum-based key metal industry chain quality collaborative improvement method

Abstract

The invention relates to the technical field of metal industry chain quality management, and discloses a key metal industry chain quality collaborative improvement method based on a quality atlas, and the method comprises the steps: collecting local quality data of each enterprise of a metal industry chain, and carrying out the preprocessing; extracting quality knowledge based on the standardized data set to preliminarily construct a quality map, and fusing and complementing the quality map; mining quality influence factor association; carrying out quality abnormity intelligent diagnosis by fusing statistical detection and map reasoning; constructing a digital twinborn simulation model, and carrying out collaborative optimization on each link of an industrial chain by adopting a hierarchical decision framework; the industrial chain quality collaborative optimization scheme is executed in stages, continuous monitoring feedback is carried out, and an optimization implementation effect report is obtained; according to the invention, through construction of the quality map covering the whole industrial chain and intelligent correlation analysis, key factors influencing the quality of the terminal product can be accurately identified, and accurate quality control is realized.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of metal industry chain quality management, more particularly, it relates to a key metal industry chain quality collaborative improvement method based on a quality map. BACKGROUND

[0002] The metal industry chain, especially the rare earth permanent magnet material industry chain, is an important foundation of modern high-tech industry and is widely used in strategic emerging industries such as new energy vehicles, wind power, aerospace, etc. However, the current metal industry chain quality management is facing serious challenges.

[0003] Taking the rare earth permanent magnet material industry chain as an example, the industry chain is usually composed of rare earth mine exploitation enterprises, rare earth separation and smelting enterprises, permanent magnet material manufacturing enterprises and downstream motor manufacturing enterprises, with many links and complex processes. The specific technical problem existing in the prior art is that the quality indicators (such as intrinsic coercivity) of the terminal permanent magnet material fluctuate too much between different batches, with a control range of ±50kA / m, while the downstream application requires control within ±30kA / m, and the quality consistency is difficult to guarantee.

[0004] The fundamental reasons for this technical problem include: quality data is scattered and lacks effective correlation, quality data of each link in the industry chain is scattered and stored in different enterprise information systems, lacking unified data standards and sharing mechanisms, and cross-link quality correlation cannot be established; quality influencing factors are complex and difficult to identify, from the composition fluctuation of upstream rare earth ore to the magnetic properties of downstream permanent magnet materials, involving dozens of process links and hundreds of influencing factors, and traditional methods are difficult to accurately identify quality influencing factors and their mutual relationship; quality abnormality tracing is difficult and time-consuming, when quality abnormalities are found, the traditional step-by-step tracing method takes 1-2 weeks, and often cannot locate the root cause, seriously affecting the quality improvement efficiency; lacking of industry chain collaborative optimization mechanism, each link usually controls quality independently, lacking of global perspective for collaborative optimization, local optimization may lead to global suboptimization, and the overall quality level of the industry chain cannot be improved; quality knowledge lacks structured representation and reuse, a large amount of quality experience and knowledge exists in unstructured form, which is difficult to effectively accumulate, share and reuse, restricting the continuous improvement of quality management level.

[0005] Therefore, there is an urgent need for a new method that can integrate industry chain quality data, intelligently identify quality influencing factors, quickly diagnose the root cause of quality abnormalities, and realize collaborative optimization decision-making, to solve the technical problems of metal industry chain quality management. SUMMARY

[0006] The present application provides a key metal industry chain quality collaborative improvement method based on a quality map, which solves the technical problems of large fluctuations in terminal product quality, difficulty in abnormality tracing, and lack of cross-enterprise collaborative optimization mechanism in related technologies.

[0007] The application provides a quality map-based key metal industry chain quality collaborative improvement method, comprising:

[0008] Collecting local quality data of each enterprise in the metal industry chain, preprocessing, and obtaining a local standardized data set;

[0009] Extracting quality knowledge based on the standardized data set to preliminarily construct a quality map, and fusing and completing the quality map to obtain a metal industry chain quality map;

[0010] Based on the metal industry chain quality map, the correlation of the quality influencing factors is mined to obtain a quality influencing factor set;

[0011] Based on the quality influencing factor set, the quality anomaly intelligent diagnosis of statistical detection and graph reasoning is fused to obtain a root cause diagnosis result;

[0012] Based on the root cause diagnosis result, the key metal industry chain quality map and the key quality influencing factor set, a digital twin simulation model is constructed, a hierarchical decision-making framework is used to collaboratively optimize each link of the industry chain, and an industry chain quality collaborative optimization scheme is obtained;

[0013] The industry chain quality collaborative optimization scheme is executed in stages and continuously monitored and fed back to obtain an optimization implementation effect report.

[0014] In a preferred embodiment, the preprocessing comprises:

[0015] Deploying an edge data acquisition node locally in each enterprise, using an industrial communication protocol to collect quality-related data in real time, and obtaining an original data set of each enterprise;

[0016] Performing data integrity checks on the original data set of each enterprise, identifying missing values, duplicate records and format errors, and using an outlier detection method based on a box plot to mark potential outliers;

[0017] Data format standardization is performed to obtain a local standardized data set.

[0018] In a preferred embodiment, the preliminary construction of the quality map based on the standardized data set comprises:

[0019] Collecting professional corpus in the metal field to form a field corpus;

[0020] Preprocessing the field corpus;

[0021] Setting up a basic framework for a field pre-training language model, self-supervised pre-training on the field corpus, and using a mask language model (MLM) task;

[0022] The pre-training process is performed on a multi-GPU server, a training round is set, the training progress is monitored according to the perplexity of the validation set, the training is stopped when the perplexity decreasing trend slows down, and a metal field pre-training language model is obtained;

[0023] Based on the metal field pre-training language model, quality entity recognition and relationship extraction tasks are performed.

[0024] In a preferred embodiment, the mining of quality influencing factor associations includes:

[0025] Based on the metal industry chain quality graph, a training task set of the meta-learning framework is constructed, and the quality influencing factor association mining problem is defined as identifying a set of upstream factor nodes from the graph and quantifying the influence degree of each factor given a target quality attribute node.

[0026] A model-independent meta-learning algorithm is used to train the meta-model for quality association mining, and the meta-model uses a heterogeneous graph attention network.

[0027] The metal industry chain quality graph is modeled using a heterogeneous graph neural network, and the embedding representation of the nodes is learned. The metal industry chain quality graph is modeled as a heterogeneous graph, and a set of node types and a set of relationship types are defined.

[0028] In a preferred embodiment, the quality anomaly intelligent diagnosis of fusion of statistical detection and graph reasoning includes:

[0029] Based on real-time quality data flow, combined with a set of quality influencing factors, a statistical process control (SPC) method is used to detect anomalies in real-time quality data. The measurement value sequence of the quality index is extracted from the historical normal production data, the mean and standard deviation are calculated, and the 3σ criterion is used to determine the control limit. When the real-time collected quality data point exceeds the control limit, it is marked as abnormal. Based on the detected set of abnormal quality indicators, combined with the metal industry chain quality graph and the set of quality influencing factors, a graph reasoning method is used for anomaly root cause tracing. Starting from the abnormal node, the causal relationship edge, process relationship edge and composition relationship edge are traced in reverse to identify the upstream cause node.

[0030] Based on the candidate cause nodes obtained by reverse tracing and the characteristics of the current anomaly, similar cases are retrieved from the historical abnormal case library to assist in root cause diagnosis. The similarity between the current abnormal subgraph and the historical case subgraph is calculated.

[0031] In a preferred embodiment, the simulation of the synergistic effect of the optimized process parameter combination of each link in the entire industry chain includes:

[0032] Based on the root cause diagnosis results and recommended solutions, combined with the metal industry chain quality map and the quality influence factor set, a digital twin simulation model of the metal industry chain is constructed, the process mechanism knowledge is extracted from the metal industry chain quality map, the mechanism simulation model of each link is constructed, and the mechanism models of each link are connected in series to form an end-to-end simulation process of the industry chain.

[0033] A hierarchical decision-making framework for industry chain quality collaborative optimization is established, and the multi-enterprise and multi-parameter collaborative optimization problem is decomposed into upper global coordination and lower local execution.

[0034] Iterative coordination and digital twin simulation verification are performed, the global coordination layer simulates the whole process from raw material input to product output based on the digital twin model, inputs the optimized process parameters of each link into the digital twin model, and predicts the quality index distribution of the end product.

[0035] In a preferred embodiment, the phased implementation of the industry chain quality collaborative optimization scheme and the continuous monitoring feedback include:

[0036] Based on the industry chain quality collaborative optimization scheme, a phased implementation plan is developed and process monitoring is performed, the implementation process is divided into several stages according to the priority, interdependence and resource constraints of each improvement task;

[0037] Based on the data collected during the implementation process, the actual effect of the industry chain quality collaborative optimization scheme is quantitatively evaluated, a baseline period and an evaluation period are set, the quality indicators before and after implementation are compared, and the quality improvement degree is calculated;

[0038] Based on the optimization implementation effect report, the quality correlation analysis model is updated online, the new data collected during the optimization implementation is added to the training data set, the graph neural network quality correlation analysis model is retrained or incrementally updated, the successful experience and failure lessons of the optimization scheme are summarized as knowledge, and the metal industry chain quality map is updated.

[0039] In a preferred embodiment, the collection of local quality data of each enterprise in the metal industry chain and the preprocessing further include:

[0040] Based on the federated learning framework, cross-enterprise data collaborative analysis is realized, the global data analysis model parameters are initialized in the central cloud server, each enterprise trains a local model copy based on the local standardized data set, the local gradient is processed with differential privacy, and the central server collects all the uploaded gradient information, then performs weighted aggregation to update the global model parameters.

[0041] In a preferred embodiment, the mining of quality influence factor correlations further includes:

[0042] To identify and quantify the impact of quality-influencing factors, a trained heterogeneous graph neural network model is used to identify key factor nodes and quantify the impact of each factor. An attention-weight-based method for identifying influencing factors is adopted. For a target quality attribute node, starting from the node, the process traces back along the edges of the graph to the upstream set of candidate influencing factor nodes. For each candidate factor node, the attention path weight of its relationship to the target node is calculated.

[0043] Using sensitivity analysis, for the identified key factor nodes, while keeping the characteristics of other factor nodes unchanged in the test samples, only the feature value of the node is changed. The change in the target quality attribute value is predicted by a graph neural network model, and the sensitivity coefficient is calculated.

[0044] By combining attention weight and sensitivity coefficient, a comprehensive influence score is defined. Based on the comprehensive score, key factors are ranked to obtain a set of quality influencing factors.

[0045] In a preferred embodiment, a computer-readable storage medium is provided for storing computer-readable instructions that, when read by a computer, enable the execution of a method for collaborative quality improvement of a key metal industry chain based on a quality map, as described above.

[0046] The beneficial effects of this invention are as follows: By constructing a multi-level quality map of the entire industry chain and using graph neural network intelligent correlation analysis technology, it achieves accurate identification and precise control of key factors affecting the quality of end products, effectively improving product quality consistency and stability; combined with multi-dimensional diagnostic technology integrating graph reasoning, statistical detection, and case matching, it enables rapid root cause localization of quality anomalies, shortens diagnostic time, and improves the accuracy of root cause diagnosis; through a hierarchical optimization framework and digital twin simulation verification technology, it enables collaborative optimization decision-making in all links of the industry chain, avoids global suboptimal problems caused by local optimization, improves overall quality level, and reduces quality costs.

[0047] By employing federated learning technology and privacy-preserving computational methods, the system enables local retention of raw data and secure sharing of model parameters among enterprises. This ensures the effectiveness of cross-enterprise collaborative data analysis while protecting corporate trade secrets, thus resolving trust issues in supply chain collaboration. Through a dynamic quality graph update mechanism and knowledge structured storage technology, the system achieves the systematic accumulation and rapid reuse of successful quality improvement experiences, avoiding repeated trial and error and improving the efficiency of quality improvement.

[0048] Based on graph-based quality impact path analysis and interpretable report generation technology, this system clearly demonstrates the causal relationships of quality issues, enabling managers to accurately understand the problem mechanisms. Its modular design architecture and standardized interface technology facilitate expansion to other metal industry chains. A real-time quality monitoring dashboard and intelligent early warning system are built to proactively prevent quality risks, fundamentally changing the traditional reactive quality management model. Attached Figure Description

[0049] Figure 1 This is a flowchart of a method for collaboratively improving the quality of a key metal industry chain based on quality maps, as described in this invention. Detailed Implementation

[0050] The subject matter described herein will now be discussed with reference to exemplary embodiments. It should be understood that these embodiments are discussed only to enable those skilled in the art to better understand and implement the subject matter described herein, and changes may be made to the function and arrangement of the elements discussed without departing from the scope of this specification. Various processes or components may be omitted, substituted, or added as needed in the examples. Furthermore, some features described in the examples may be combined in other examples.

[0051] At least one embodiment of the present invention discloses a method for synergistic quality improvement of a key metal industry chain based on quality maps, such as... Figure 1 As shown, it includes:

[0052] Step 1: Collect local quality data from various enterprises in the metal industry chain, perform preprocessing, and obtain a local standardized dataset;

[0053] Step 1.1: Deploy edge data acquisition nodes and acquire raw multi-source data. Based on the existing manufacturing execution systems, quality management systems, testing equipment, industrial sensors, and other data sources of each enterprise in the industry chain, deploy edge data acquisition nodes locally in each enterprise. Use industrial communication protocols such as OPC UA and MQTT to collect quality-related data in real time, obtaining the raw data sets of each enterprise. Collect key quality data from each link of the industry chain, including multi-dimensional data such as raw material composition, process parameters, equipment status, intermediate product quality indicators, and final product performance parameters. Through the above multi-dimensional and multi-level data acquisition, obtain the raw data sets of each enterprise. Each dataset contains time-series structured data, semi-structured inspection reports, and unstructured process documents and anomaly records.

[0054] Step 1.2, Local Data Preprocessing and Quality Cleaning. Based on the computing power of each enterprise's edge nodes, the original datasets of each enterprise are subjected to data integrity checks, identifying missing values, duplicate records, and format errors. An outlier detection method based on box plots is used. For marked outliers, a secondary judgment is made combining time series characteristics and physical constraints. If the outlier is confirmed to be caused by measurement errors or equipment failures, it is removed or imputed using linear interpolation based on adjacent time data. If it is a genuine process anomaly, it is retained and marked as an anomalous event. For missing data, a multiple imputation method is used to fill in the missing data based on historical statistical characteristics of similar equipment or batches. For missing quality indicators, if they cannot be reliably imputed, they are marked as missing and handled specially in subsequent analysis. Data format standardization is performed, converting timestamps exported from different systems to UTC time format and different units of measurement to the International System of Units (SI). For categorical data, a unified coding system is established for equipment numbers, material batch numbers, process types, etc. For unstructured text data, natural language processing technology is used for preliminary parsing, extracting key entities and numerical information, and converting it into a semi-structured format. After the above processing, a standardized local dataset is obtained for each enterprise. This dataset is kept locally by the enterprise and is not uploaded to an external server to ensure data privacy and security.

[0055] Step 1.3: Training a cross-enterprise collaborative data analysis model based on federated learning. This step utilizes a federated learning framework to achieve cross-enterprise collaborative data analysis while protecting the data privacy of each enterprise. Global data analysis model parameters are initialized on a central cloud server. This model is a deep neural network structure used to learn the feature representation of quality data and the predictive relationship of quality indicators. The initialized model parameters are distributed to the n participating enterprises in the industry chain. Each enterprise trains a local copy of the model based on its local standardized dataset. Specifically, the i-th enterprise uses local data to perform k rounds of mini-batch gradient descent training to minimize the local loss function. The mean squared error is used to measure the deviation between the predicted and actual values ​​of the quality indicators, and the gradient of the local model parameters is calculated.

[0056] To protect data privacy, each enterprise performs differential privacy processing on its local gradients by adding Laplace or Gaussian noise, resulting in a noisy gradient. The scale of the noise is determined based on the privacy budget ε and sensitivity, with ε ranging from 0.5 to 2. Each enterprise uploads the noisy gradient, rather than the original data, to a central server. The central server cannot deduce the enterprise's original data from the gradient information, thus achieving privacy protection.

[0057] After the central server collects gradient information uploaded by all n companies, it performs weighted aggregation to update the global model parameters, using the FedAvg algorithm to ensure that companies with larger datasets contribute more to the global model. The updated global model parameters are then redistributed to the companies for the next round of federated learning iterations. The global parameter update process involves subtracting the learning rate multiplied by the weighted average of the gradients from all companies, including the noisy gradients, from the current round's global parameters, where the weights are the aggregate weights for each company.

[0058] The process of local training, gradient uploading, global aggregation, and model deployment is repeated until the global model converges. The convergence criterion is that the decrease in the global loss function is less than a preset threshold or the maximum number of iterations is reached in a series of consecutive iterations. After federated learning training, a converged global collaborative data analysis model is obtained. This model integrates the local data characteristics of each enterprise in the industry chain and has the ability to comprehensively analyze the quality data of the entire chain. At the same time, the original data of each enterprise is always kept locally to meet privacy protection requirements.

[0059] This step outputs the local standardized datasets for each enterprise and the global data analysis model trained by federated learning.

[0060] Step 2: Extract quality knowledge based on standardized datasets to initially construct a quality map, and then fuse and complete the quality map to obtain a quality map of the metal industry chain;

[0061] Step 2.1: Construct a pre-trained language model for the metal domain. Based on the local standardized dataset output in Step 1, and combined with a large-scale corpus of collected metal domain data, a pre-trained language model for the domain is constructed. Professional corpus data in the metal domain is collected, including academic papers, technical standards, patent documents, process manuals, testing report templates, industry white papers, and other text data related to rare earth, lithium, cobalt, and other metals, forming a domain corpus exceeding 10GB in size, covering millions of professional terms and technical expressions. The domain corpus undergoes preprocessing, including text cleaning, sentence segmentation, word segmentation, and vocabulary construction. A pre-trained language model for the metal domain is constructed using the BERT (Bidirectional Encoder Representations from Transformers) architecture, learning the semantics of domain professional terms and knowledge representation patterns through a masked language model task. This pre-trained model can generate high-quality contextual embedding representations of domain vocabulary and sentences, laying the foundation for subsequent high-quality entity recognition and relation extraction tasks.

[0062] Step 2.2: Quality Entity and Relationship Labeling Based on Active Learning. Based on a pre-trained language model in the metal domain, fine-tuning training is performed for quality entity recognition and relation extraction tasks, employing an active learning strategy to improve labeling efficiency. The entity type set of the quality map is defined, including material entities, element entities, process entities, equipment entities, quality attribute entities, and numerical entities; the relation type set is defined, including compositional relations, process relations, inspection relations, causal relations, and traceability relations. An initial labeling set is randomly sampled from textual data such as technical documents, inspection reports, and process specifications from various companies in the industry chain. This initial set contains 100 to 200 documents, which are manually labeled by domain experts, annotating the quality entities appearing in each document and their types, as well as the relationships and types between entities, resulting in the initial labeled dataset. Based on the initial labeled dataset, the pre-trained language model in the metal domain is fine-tuned to train a Named Entity Recognition (NER) model and a Relation Extraction (RE) model.

[0063] After obtaining the initial NER and RE models, an active learning loop is initiated to efficiently augment the labeled data. From a large collection of unlabeled documents, the current model is used to perform entity recognition and relation extraction predictions for each document, calculating the prediction uncertainty for each document. The K documents with the highest uncertainty are selected and submitted to domain experts for annotation, where K is set to 20 to 50. The annotated documents are added to the labeled dataset, and the updated labeled dataset is used to fine-tune the NER and RE models. This active learning loop is repeated 10 to 20 times, with each iteration selecting the samples with the highest uncertainty for annotation. This ensures that the annotation work focuses on the most uncertain and informative difficult samples, reducing the number of labeled samples required to achieve the target performance.

[0064] Step 2.3: Extract quality knowledge from multi-source data and construct a quality graph. Based on the trained NER and RE models, quality knowledge is extracted from the local standardized dataset obtained in Step 1 and technical documents to construct a quality graph. For structured numerical quality data, a rule-based mapping method is used to directly convert it into graph entities and relationships. For example, records of a batch of rare earth ore are extracted from the production database, including batch number Batch001, REO content 65.3%, Nd content 22.1%, Dy content 0.8%, Fe impurity 0.15%, etc. Based on this, a material entity node "Rare Earth Ore_Batch_001" is created, and element entity nodes "REO", "Nd", "Dy", and "Fe" are created. Composition relationship edges are created between the material entities and each element entity, with the attributes of the edges being the content value and measurement time. Similarly, process data is extracted to create process entity nodes and process parameter attributes, and test data is extracted to create quality attribute nodes and numerical attributes.

[0065] For semi-structured and unstructured text data, NER and RE models are used for knowledge extraction. The NER model identifies entities, and the RE model identifies relationships.

[0066] Knowledge extracted from all data sources is integrated and stored in a graph database. Graph database systems such as Neo4j or JanusGraph are used, supporting efficient graph structure storage and graph query operations. In the database, node labels (Labels) correspond to entity types, and relation types (RelationTypes) correspond to relation types. Each node and edge has an attribute dictionary storing specific information. After knowledge extraction and integration, a preliminary quality graph is obtained, containing a set of nodes, edges, and attributes. For applications in the rare earth permanent magnet material industry chain, the number of graph nodes reaches 100,000, including thousands of material batch nodes, hundreds of process nodes, hundreds of equipment nodes, and dozens of quality attribute nodes; the number of edges reaches 500,000, including a large number of compositional relationships, process relationships, causal relationships, and traceability relationships.

[0067] Step 2.4, Knowledge Fusion and Completion of the Quality Graph. Based on the preliminary quality graph, knowledge fusion, entity alignment, and knowledge completion are performed to improve the quality and completeness of the quality graph. Entity alignment is performed to identify and merge duplicate entity nodes pointing to the same real object. Different data sources and documents may use different names or abbreviations to refer to the same entity. For example, "neodymium iron boron," "NdFeB," and "neodymium iron boron alloy" refer to the same type of material, and "coercivity" and "intrinsic coercivity" refer to the same quality attribute. An entity alignment method based on string similarity and semantic embedding similarity is used to calculate the similarity between candidate entity pairs. Entity pairs with a similarity greater than 0.85 are determined to be the same entity and merged. String similarity is calculated using edit distance or Jaccard similarity, and semantic embedding similarity is calculated by using a pre-trained model to generate vector representations of entity names and then calculating cosine similarity. For multiple nodes determined to be the same entity, their attribute information is merged, retaining more complete and up-to-date attribute values, updating the connections of all related edges, and deleting redundant nodes.

[0068] Perform relation validation and filter out erroneous relation edges. Validate extracted relations based on domain rules and statistical confidence scores. Domain rules include: component relations can only exist between material entities and element entities; process relations must start with a material and end with a process; causal relations must conform to chronological order, with cause preceding effect. Check whether all relation edges in the quality graph satisfy the corresponding rule constraints; edges that do not satisfy are marked as suspicious. Statistical confidence scores are based on the frequency of relation occurrence in the training data and the model's prediction probability; relation edges with a prediction probability below 0.6 in the RE model are marked as low-confidence relations. Submit suspicious and low-confidence relations to domain experts for manual review. Relationships confirmed as erroneous by experts are deleted, while correct relationships are retained and their weights increased.

[0069] Knowledge completion is performed to supplement missing relationships and attributes in the quality graph. A link prediction method based on graph embedding is used to learn low-dimensional vector representations of entities and relationships in the quality graph. A knowledge graph embedding model is used to map each entity to a d-dimensional vector, where d typically takes the value 100 to 300, and each relationship type is also mapped to a d-dimensional vector or transformation matrix. The training objective is to ensure that entity pairs (head, tail) with relationship r satisfy a certain vector space constraint, such as the TransE model requiring that head plus the relation vector is approximately equal to the tail vector. The embedding model is trained on existing triples (entity 1, relation, entity 2), so that positive sample triples have high score function values ​​and negative sample triples have low score function values. After training, the score is calculated for all possible combinations of entity pairs and relationship types in the graph. High-scoring triple candidate relationships may exist but are missing in the current graph. High-scoring candidate relations are submitted to experts for review and then added to the graph. For missing entity attributes, collaborative filtering or attribute inference methods based on similar entities are used to fill in the missing attributes. For example, if a batch of materials is missing a certain quality attribute value, other batches in the spectrum that are most similar to it in terms of composition and process are found, and the mean of their attribute values ​​is used as the estimated value to fill in the missing attribute value.

[0070] A time-series expansion mechanism for the quality graph is established, adding time dimension information to nodes and edges. A creation time attribute is added to each node, corresponding to the production or inspection time of the material batch; an occurrence time attribute is added to each edge, corresponding to the process handling time or the occurrence time of the quality event. Based on timestamp information, a snapshot of the graph's state at any historical moment can be queried, supporting time-series analysis and trend prediction of quality data. A time-series index is constructed to accelerate graph queries based on time ranges. Through the above knowledge integration and completion, a quality graph of the metal industry chain is obtained.

[0071] This step outputs a quality map of the metal industry chain, containing 100,000 nodes and 500,000 relationship edges. Node types cover materials, processes, equipment, and quality attributes, while relationship types cover composition, processes, causality, and traceability. Nodes and edges are equipped with rich attribute information and timestamps. Simultaneously, it outputs a continuously updated domain-pre-trained model and fine-tuned NER and RE models for subsequent incremental knowledge extraction.

[0072] Step 3: Based on the quality map of the metal industry chain, explore the correlation of quality influencing factors to obtain a set of quality influencing factors;

[0073] Step 3.1: Construct a meta-learning training task set for quality correlation mining. Based on the metal industry chain quality map output in Step 2, construct a training task set for the meta-learning framework to enable the model to quickly learn new quality correlations in small sample scenarios. The problem of quality influencing factor correlation mining is defined as follows: Given a target quality attribute node, such as the intrinsic coercivity of permanent magnet materials; identify a set of upstream factor nodes that have a certain influence on this quality attribute from the map, such as raw material composition and process parameters; and quantify the degree of influence of each factor. Multiple subgraph tasks are constructed by sampling from the quality map, each subgraph task corresponding to a specific quality attribute correlation mining target. The sampling strategy is as follows: select a quality attribute node as the task target node, start from this node, and trace back K-hop neighborhoods along causal relationship edges, process relationship edges, and composition relationship edges in the map, where K takes values ​​from 2 to 4, extracting a subgraph containing the target node and all nodes and edges within its K-hop neighborhood. For each subgraph task, multiple batches of samples for that quality attribute are extracted from historical data. Each sample contains the attribute value of the upstream factor node and the actual measured value of the target quality attribute, where the attribute value of the upstream factor node serves as the input feature and the actual measured value of the target quality attribute serves as the supervision label. The samples for each subgraph task are randomly divided into a support set and a query set. The support set contains 5 to 20 samples to simulate a small sample scenario, while the query set contains 50 to 100 test samples. Through this sampling process, a meta-training task set is constructed, containing M tasks, where M ranges from 50 to 200, covering different quality attributes, different material types, and different process paths to ensure task diversity. The construction of the meta-training task set ensures that the model has encountered multiple types of quality-related learning tasks during the training phase, thus possessing the meta-learning capability to quickly adapt to new tasks.

[0074] Step 3.2, Training the Meta-Model Based on MAML. Based on the meta-training task set, a meta-model for quality association mining is trained using the Model-Agnostic Meta-Learning (MAML) algorithm. The meta-model employs a graph neural network architecture, specifically the HetGAT heterogeneous graph attention network, capable of handling heterogeneous graph structures with multiple node types and relationship types in the quality graph. The HetGAT model contains L graph convolutional layers, where L ranges from 2 to 4. Each layer uses a type-specific attention mechanism to aggregate information for different types of nodes and relationships. The model input consists of the initial feature vectors of each node in the subgraph. For material nodes, the features include their component content vectors; for process nodes, the features include process parameter vectors; and for quality attribute nodes, the features include historical statistical values. The model output is the embedding vector of the target quality attribute node, which is then passed through a fully connected regression layer to predict the numerical value of the target quality attribute.

[0075] The MAML training process consists of two layers: an inner loop and an outer loop. The meta-model parameters are initialized. In each iteration of the outer loop, a batch of B tasks is randomly sampled from the meta-training task set, where B ranges from 4 to 16. For each task in the batch, inner loop adaptation is performed: the model is trained using the support set of that task, and a loss function on the support set is calculated. This loss is expressed as mean squared error (MSE), measuring the deviation between the model's predicted quality attribute values ​​and the true values. Inner loop adaptation allows the model parameters to be quickly adjusted to adapt to the specific quality association patterns of the current task.

[0076] After the inner loop adaptation is completed, the performance of the adapted model is evaluated on the query set of the task, and the query set loss is calculated. The query set loss is calculated as follows: after the inner loop adaptation of MAML meta-learning is completed, task-specific parameters trained and adjusted on the support set are used to input each sample in the query set into the adapted model for prediction. The sum of the squared prediction errors of all samples in the query set is divided by the number of samples to obtain the query set loss for the task.

[0077] The goal of the outer loop is to optimize the meta-model parameters so that the model adapted by the inner loop performs optimally on the query set, i.e., minimizing the expected value of the query set loss for all tasks. This outer loop iteration process is repeated several times (1000 to 5000 times in this example) until the average query set loss of the meta-model on the validation task set converges and no longer decreases, yielding the trained meta-model parameters. The trained meta-model possesses rapid learning capabilities. When faced with new quality association mining tasks, it only needs a few steps of gradient descent adaptation using a small number of support set samples to achieve high prediction accuracy on the query set, overcoming the learning difficulties in small sample scenarios.

[0078] Step 3.3, Heterogeneous Graph Neural Network Modeling and Node Embedding Learning. Based on the meta-model and transfer learning model obtained in Steps 3.2 and 3.3, heterogeneous graph neural network modeling is performed on the metal industry chain quality map to learn the embedded representation of nodes. The metal industry chain quality map is modeled as a heterogeneous graph, defining the node type set as {material, element, process, equipment, quality attribute, numerical}, and defining the relationship type set as {composition, process treatment, detection, causality, traceability, parameter}. For different types of nodes in the graph, type-specific initial feature vectors are constructed. The features of material nodes include the content vector of their key component elements and the unique thermal encoding of the material category; the features of element nodes include the physicochemical properties of the element such as atomic number, atomic weight, and electronegativity; the features of process nodes include the embedding vector of the process type and the numerical vector of key process parameters; the features of equipment nodes include the equipment type and equipment capability level; the features of quality attribute nodes include the embedding vector of the attribute name and historical statistical values; the features of numerical nodes are standardized numerical values.

[0079] A heterogeneous graph attention network, HetGAT, is used for multi-layer information propagation and node embedding updates. In the l-th layer, for the i-th node, its embedding vector is updated by aggregating information from its neighboring nodes. Since neighboring nodes may have different types, a type-specific attention mechanism is adopted. For each relationship type, a relationship-specific attention weight calculation function and message passing function are defined. The importance weights between nodes are calculated using the attention mechanism, and the feature representation of each node is updated by weighted aggregation of neighboring node information. Through graph convolutional propagation of L layers (L is 3 in this embodiment), the embedding vector of each node fuses the structural and semantic information within its L-hop neighborhood. The node embedding dimension is set to 256 or 512 dimensions.

[0080] For a target quality attribute node, its embedding vector, obtained after L-layer graph convolution, contains aggregated information from all upstream nodes reachable from that node, implicitly encoding the influence of upstream material composition, process parameters, equipment status, and other factors on intrinsic coercivity. A graph neural network is trained on historical batch data, with the optimization objective being to enable the node embeddings to accurately predict quality attribute values. Specifically, the embedding vectors of the target quality attribute nodes are input into a multilayer perceptron regressor, which outputs the predicted intrinsic coercivity value. The Adam optimizer is used for training, with a learning rate of 0.001 and a batch size of 32, for several epochs until the validation set loss no longer decreases. After training, the model can accurately predict downstream quality attributes based on the input features of upstream factors, and the learned node embedding vectors and attention weights imply the correlation between quality influences.

[0081] Step 3.4: Identify and quantify the impact factors on quality. Based on the heterogeneous graph neural network model trained in Step 3.4, identify key factor nodes that have a certain impact on the target quality index and quantify the impact of each factor. An attention-weight-based method for identifying impact factors is adopted. For a target quality attribute node, starting from that node, trace back upstream along the edges of the graph to the set of candidate impact factor nodes. This set includes upstream material nodes, process nodes, process parameter nodes, equipment nodes, etc. For each candidate factor node, calculate its attention path weight to the target node. The attention path weight is defined as the aggregation of attention weights on all paths from the candidate factor node to the target node. The specific calculation method is as follows: Through the forward propagation of the graph neural network, the attention weight of each edge in each layer is recorded. Starting from the candidate node, depth-first search or breadth-first search is used to find all paths to the target node, with the path length limited to K hops (e.g., K equals 3). For each path, the path weight is calculated as the product of the attention weights of all edges on the path. The total influence weight of a candidate node is equal to the sum or maximum value of all path weights from that node to the target node. The specific choice between sum or maximum depends on the path semantics; here, the maximum value represents the strongest influencing path. After calculating the influence weights of all candidate factor nodes, they are sorted from largest to smallest weight, and the top K nodes with the highest weights are selected as the set of key influencing factors, where K ranges from 10 to 30.

[0082] To further quantify the impact of key influencing factors on the target quality indicators, sensitivity analysis was employed. For identified key factor nodes, while keeping the characteristics of other factor nodes unchanged in the test sample, only the characteristic value of this node was changed, for example, by increasing or decreasing it by 10%. A graph neural network model was used to predict the change in the target quality attribute value. The sensitivity coefficient was calculated as follows: the change in the target quality indicator was divided by its original value to obtain the relative change; the change in the influencing factor was divided by its original value to obtain the relative change of the factor; the relative change in the target quality indicator was divided by the relative change of the influencing factor to obtain the sensitivity coefficient. The larger the absolute value of the sensitivity coefficient, the more significant the impact of the factor on the target quality indicator. Combining attention weight and sensitivity coefficient, the comprehensive influence score was calculated as follows: the absolute values ​​of the attention weight and sensitivity coefficient were normalized; the normalized attention weight was multiplied by a weight coefficient, and the absolute value of the normalized sensitivity coefficient was multiplied by another weight coefficient, with the sum of the two weight coefficients being 1; the two were then added to obtain the comprehensive influence score. Based on the comprehensive score, the key factors were ranked to obtain the final set of quality influencing factors and their quantified influence.

[0083] Step 3.5: Extract the quality impact transmission path and generate an interpretability report. Based on the key influencing factors identified in Step 3.5, graph path analysis and interpretability methods are used to extract the quality impact transmission path and generate an interpretability report to help domain experts understand and verify the quality correlations discovered by the model. For each key influencing factor node, the strongest influence path from the key influencing factor node to the target quality attribute node is extracted. The strongest influence path is defined as the path with the largest attention path weight, which represents the main transmission mechanism of quality impact. An attention weight-based path search algorithm is used: starting from the key influencing factor node, at each step, the outgoing edge with the largest attention weight is selected for expansion until the target node is reached or the maximum path length limit is reached. The path obtained by the search is the strongest influence path. Semantic interpretation is performed on each node and edge in the path. For example, the path "Rare earth ore Dy content → Batching Dy ratio → Sintering process magnetic field → Grain boundary phase Dy enrichment → Intrinsic coercivity" describes the complete causal chain of how the Dy element content in the raw material affects the macroscopic magnetic property intrinsic coercivity through batch ratio control, sintering process regulation, and microstructure evolution.

[0084] The GNNExplainer interpretability method is used to generate causal explanatory subgraphs. GNNExplainer uses an optimization method to find the subgraph structure that contributes most to the model's prediction results. The optimization objective is to maximize the prediction probability of the mask image minus the subgraph complexity penalty, and gradient ascent is used to optimize the mask parameters. The resulting subgraph after optimization is the causal explanatory subgraph, which contains the nodes and edges most critical to the prediction of the target quality attributes. The explanatory subgraph is visualized, with nodes colored by type and edges displayed according to attention weight, and the attribute values ​​of key nodes and the relationship types of key edges are labeled.

[0085] Generate a quality influencing factor correlation analysis report, which includes: a list of key influencing factors, listing the names, types, influence weights, sensitivity coefficients, and comprehensive scores of the top-K influencing factor nodes, presented in tabular form; a quality influence transmission path diagram, showing the strongest path from each key factor to the target quality attribute, presenting the node sequence and relationship types in the path in flowchart form; a causal explanation subgraph, visually displaying the subgraph structure containing key influence mechanisms; a sensitivity analysis chart, plotting the curves of the impact of key factor changes on the target quality indicator, with the horizontal axis representing the percentage change of the factor and the vertical axis representing the percentage change of the quality indicator; and expert validation suggestions, providing checkpoints and suggestions from domain experts for validation of the key factors and influence paths identified by the model.

[0086] This step outputs a set of quality-influencing factors, including a set of quality influence transmission paths, a causal explanation subgraph, a correlation analysis report of quality-influencing factors, a trained heterogeneous graph neural network model, and a meta-learning model.

[0087] Step 4: Based on the set of quality influencing factors, intelligent diagnosis of quality anomalies is carried out by integrating statistical detection and graph reasoning to obtain the root cause diagnosis results;

[0088] Step 4.1, Real-time Quality Anomaly Detection Based on Statistical Process Control. Based on the real-time quality data stream collected in Step 1, and combined with the set of quality influencing factors output in Step 3, the Statistical Process Control (SPC) method is used to detect anomalies in the real-time quality data. A real-time monitoring mechanism is established for quality indicators at each stage of the industrial chain. Taking the intrinsic coercivity of rare earth permanent magnet materials as an example, the measurement sequence of intrinsic coercivity is extracted from historical normal production data excluding batches marked as abnormal. The mean and standard deviation of the intrinsic coercivity are calculated, and the control limits are determined using the 3σ criterion. The upper control limit equals the mean of the intrinsic coercivity plus three times the standard deviation of the intrinsic coercivity, and the lower control limit equals the mean of the intrinsic coercivity minus three times the standard deviation of the intrinsic coercivity. Similarly, control limits are established for other quality indicators and key influencing factors.

[0089] After a new production batch completes quality inspection, the real-time collected quality data points are entered into the monitoring system. For each quality indicator's measured value, it is determined whether it exceeds the control limit: if the measured value of a quality indicator is greater than the upper control limit or less than the lower control limit, the indicator is marked as abnormal, triggering an anomaly warning. In addition to single-point exceedance judgment, the run test rule is also used to judge process stability, including: 7 consecutive points located on the same side of the center line, 6 consecutive points monotonically rising or falling, 14 consecutive points alternating up and down, etc. If these patterns are detected, the process is judged as abnormal. For joint monitoring of multiple quality indicators, the multivariate statistical process control (MSPC) method is used to calculate the Hotelling T² statistic, comprehensively considering the joint distribution of multiple indicators. T² is equal to the transpose of the sample deviation vector multiplied by the inverse of the covariance matrix multiplied by the sample deviation vector. When T² exceeds the critical value of the chi-square distribution, it is judged as a multivariate anomaly. Through the above statistical detection methods, abnormal signals in the quality data are identified in real time, resulting in a set of abnormal quality indicators. For each abnormal indicator, its measured value, deviation degree, detection time, and other information are recorded.

[0090] Step 4.2, Anomaly Root Cause Tracing Based on Graph Reasoning. Based on the set of abnormal quality indicators detected in Step 4.1, combined with the metal industry chain quality graph constructed in Step 2 and the set of quality influencing factors identified in Step 3, a graph reasoning method is used to trace the root causes of anomalies. For the detected abnormal quality indicators, i.e., quality indicators that deviate from the normal range, the corresponding quality attribute node is located in the metal industry chain quality graph. Starting from the abnormal node, tracing backwards along causal relationship edges, process relationship edges, and composition relationship edges, all upstream cause nodes that may have caused the anomaly are identified. The backward tracing uses a reverse breadth-first search algorithm of the graph. Starting from the abnormal node, all incoming edges (i.e., edges ending at the abnormal node) are traversed. The starting nodes of the incoming edges are added to the queue as candidate cause nodes. This process is repeated until the tracing depth reaches the preset maximum number of hops or all reachable nodes are traversed. The maximum number of hops is 3 to 5 hops. Through backward tracing, a set of candidate cause nodes is obtained, which includes various types of nodes such as upstream material batch nodes, process nodes, process parameter nodes, and equipment nodes.

[0091] For each node in the candidate cause node set, check whether its state is abnormal within the time window of the anomaly occurrence. Extract the attribute values ​​of the node during the abnormal batch production period from the time-series data of the metal industry chain quality map, and compare them with the normal range of that node. The normal range is determined by the statistical distribution of historical normal batch data, using the mean ± 2 standard deviations or the mean ± 3 standard deviations as the normal range. If the current attribute value of a node exceeds the normal range, the node is considered to have an abnormal state, and its probability as the root cause increases. Record the abnormal state identifier and abnormality degree of each candidate cause node, where the abnormal state identifier is 1 to indicate abnormality and 0 to indicate normality, and the abnormality degree refers to the deviation factor of the current value from the center of the normal range.

[0092] Combining the quality influencing factor set and influence weight information output in step 3, priority weights are applied to candidate cause nodes. If a candidate cause node is also a previously identified key influencing factor, its priority weight is increased. Considering the existence of the causal path, the anomalousness of the node's state, and the node's importance among the key factors, a graph inference probability score is calculated for the candidate cause node. The indicator function is 1 if the node belongs to the key factor set, and 0 otherwise. The path length is the shortest path length from the candidate cause node to the anomalous node, and the key factor weighting coefficient ranges from 0.5 to 2. A higher graph inference probability score indicates that the node is more likely to be the root cause of the anomaly.

[0093] The specific steps for calculating the probability score of graph reasoning are as follows: multiply the abnormal state identifier of the candidate cause node by the deviation factor, then multiply by the key factor weight adjustment term, and divide by the shortest path length from the node to the abnormal node to obtain the probability score of graph reasoning.

[0094] Step 4.3, Historical Similar Anomaly Retrieval Based on Case Matching. Based on the candidate cause nodes obtained from the reverse tracing in Step 4.2 and the features of the current anomaly, similar cases are retrieved from the historical anomaly case database to assist in root cause diagnosis. The historical anomaly case database stores past quality anomaly events and their diagnostic results. Each case record includes: anomaly quality indicator name, anomaly value, anomaly time, involved material batch, upstream cause node, root cause, solution, case subgraph, etc. For the current anomaly, its feature vector is extracted as a query. Features include: type of anomaly quality indicator, deviation direction, deviation degree, involved material type, process path, and time characteristics of anomaly occurrence. A subgraph containing key information is extracted from the anomaly node and its candidate cause nodes. This subgraph contains the anomaly node, candidate cause nodes, and the critical path between them.

[0095] In the historical case database, each case corresponds to a case subgraph. The similarity between the current abnormal subgraph and the historical case subgraphs is calculated. Graph similarity is calculated using either the graph kernel method or graph edit distance. Graph edit distance is defined as the minimum number of edit operations required to transform one graph into another. Edit operations include node insertion, node deletion, node replacement, edge insertion, edge deletion, and edge replacement, each with a corresponding cost. For quality graph subgraphs, the node replacement cost is defined to be related to the differences in node type and attributes; replacement costs are low for nodes of the same type and high for nodes of different types. Edge replacement costs are related to the differences in relation types. The A* algorithm or dynamic programming algorithm is used to calculate the graph edit distance; the smaller the distance, the more similar the two subgraphs are.

[0096] In addition to subgraph structural similarity, semantic similarity of feature vectors is also calculated. Feature vectors of current anomalies and historical cases are encoded, including discrete and continuous features such as anomaly indicators, deviation levels, and material types. Discrete features are encoded using one-hot encoding or embedded vectors, while continuous features are normalized and concatenated into a feature vector. The cosine similarity of the feature vectors is calculated. Combining subgraph similarity and feature similarity, the overall case similarity is calculated by multiplying the structural similarity weight by the subgraph similarity, and then adding the feature similarity weight multiplied by the feature similarity. The overall similarity is calculated for all cases in the historical case database, and the top N cases with the highest similarity are selected as the set of similar cases, where N ranges from 3 to 10.

[0097] Root cause information is extracted from similar cases. For each similar case, the root cause nodes and solutions from historical diagnoses are recorded. The frequency of each candidate cause node in the top N similar cases is statistically analyzed, and the frequency is normalized to a probability, yielding a probability based on case matching. This probability reflects the indicative role of historical experience in identifying the current abnormal root causes.

[0098] Step 4.4, Root Cause Diagnosis Decision and Impact Assessment Based on Multi-Source Evidence. Based on the probability score of the graph inference from Step 4.2, the case matching probability from Step 4.3, and the quality influencing factor set and influence weight information output from Step 3, candidate cause nodes are comprehensively scored to obtain the final root cause diagnosis result. For candidate cause nodes, the data-based anomaly probability also needs to be calculated using the Z-score. The calculation steps are as follows: subtract the historical mean of the node's attribute from the current attribute value of the candidate cause node, divide by the historical standard deviation to obtain the Z-score; input the absolute value of the Z-score into the standard normal cumulative probability function to obtain the data-based anomaly probability. The larger the absolute value of the Z-score, the more anomalous the current value, and the closer the data-based anomaly probability is to 1.

[0099] Based on the evidence from the three dimensions mentioned above, a comprehensive score is calculated for each candidate cause node, where the sum of the three weighting coefficients is 1. These coefficients can be adjusted according to the reliability of each evidence source in practical applications, for example, to 0.4, 0.3, and 0.3 respectively. All candidate cause nodes are then sorted from highest to lowest comprehensive score to obtain a root cause diagnosis ranking list. The top 3 to 5 nodes with the highest comprehensive scores in the ranking list are the most likely root causes, thus yielding the root cause diagnosis result.

[0100] Based on the diagnosed root cause, the impact of the anomaly on downstream processes is assessed. Starting from the anomalous quality attribute node, the process propagates forward along the traceability and process relationship edges in the quality graph to identify all downstream batches and end products that used the anomalous batch of materials. Specifically, the graph queries the set of all downstream material batch nodes that used the current anomalous batch as raw material. For each downstream batch node, its current status is checked: if the batch has been produced but not yet delivered, it is marked as a potentially risky batch requiring quality re-inspection; if the batch has been delivered to downstream customers, it is marked as a batch requiring early warning notification. A breadth-first search of the graph is used to traverse all downstream nodes until the end of the supply chain is reached or the traversal depth limit is reached. The number of affected downstream batches, the number of customers involved, and potential economic losses are statistically analyzed to generate an impact assessment report. This report guides quality management personnel to take targeted risk control measures, such as isolating and testing risky batches, issuing quality warnings for delivered batches, and recalling anomalous batches.

[0101] This step outputs the root cause diagnosis results, including the 3 to 5 most likely root cause nodes, their composite scores and sources of evidence; abnormal quality indicators and their measurements and deviations; an impact scope assessment report, including a list of affected downstream batches and their risk levels; an abnormal diagnosis subgraph, visually displaying abnormal nodes, root cause nodes and their causal paths; and recommended solutions.

[0102] Step 5: Based on the root cause diagnosis results, the quality map of the key metal industry chain, and the set of key quality influencing factors, construct a digital twin simulation model, and use a hierarchical decision-making framework to collaboratively optimize each link of the industry chain to obtain a collaborative optimization scheme for the quality of the industry chain.

[0103] Step 5.1: Construct a digital twin simulation model of the industrial chain. Based on the root cause diagnosis results output in Step 4 and the recommended solutions extracted from similar historical cases, combined with the metal industrial chain quality map constructed in Step 2 and the set of quality influencing factors identified in Step 3, a digital twin simulation model of the metal industrial chain is constructed to provide a virtual testing environment for verifying targeted collaborative optimization solutions. The digital twin model consists of three layers: a physical layer, a data layer, and a model layer. The physical layer corresponds to the actual physical entities of each link in the industrial chain, including mines, smelters, manufacturing workshops, equipment, etc.; the data layer corresponds to the multi-source real-time data and historical data collected in Step 1; and the model layer contains a simulation model based on mechanistic knowledge and data-driven approaches.

[0104] Process mechanism knowledge is extracted from the quality map of the metal industry chain to construct mechanism simulation models for each link. For rare earth extraction and separation processes, a mathematical model of the extraction process is established based on extraction equilibrium theory and mass transfer kinetic equations. The model inputs are the rare earth element concentration vector of the feed liquid, the type and concentration of the extractant, the extraction temperature, the phase ratio, and the number of extraction stages. The model outputs are the rare earth element concentration distribution in each extraction stage and the purity of the final product. The partition ratio calculation steps in the extraction equilibrium equation are as follows: the target element concentration in the organic phase is divided by the target element concentration in the aqueous phase to obtain the partition ratio. The partition ratio is related to factors such as extractant concentration, pH value, and temperature, and is calculated using empirical correlation equations or thermodynamic models. The multi-stage countercurrent extraction process uses the inter-stage material balance equations for iterative solution. The aqueous phase discharge concentration of stage j is equal to the aqueous phase feed concentration of stage j+1 minus the amount extracted into the organic phase. The organic phase discharge concentration is equal to the organic phase feed concentration of stage j-1 plus the amount extracted from the aqueous phase. The concentration distribution of each stage is obtained by iterative calculation using the partition ratio relationship. For the sintering process of permanent magnet materials, a simulation model of the sintering process is established based on phase transformation kinetics and grain growth theory. The model inputs include powder formulation, sintering temperature profile, holding time, and sintering atmosphere. The model outputs the sintered density, grain size, grain boundary phase composition, and macroscopic magnetic properties. The densification process during sintering is described by a diffusion equation. The rate of change of relative density over time is related to the diffusion coefficient and temperature. The diffusion coefficient is calculated using the Arrhenius equation. The relationship between magnetic properties and microstructure is established using a micromagnetic model or empirical formula. Coercivity is related to grain size and grain boundary phase composition. The calculation steps are: subtract the fitting coefficient b multiplied by the grain size from the fitting coefficient a, and then add the fitting coefficient c multiplied by the Dy element content in the grain boundary phase.

[0105] By linking and combining the mechanistic models of each stage, an end-to-end simulation process of the industrial chain is constructed. The input is the composition and mining volume of the upstream raw ore. After passing through the separation and smelting model, the composition and purity of the intermediate product are obtained. Then, the input is fed into the permanent magnet material preparation model to obtain the magnetic properties of the final product, realizing the simulation of the entire process from raw materials to products. For complex processes that cannot be accurately described by the mechanistic model, a data-driven surrogate model is used to supplement it. Based on the historical data of steps 1 and 3, a neural network or Gaussian process regression model is trained to establish a black-box mapping relationship from input process parameters to output quality indicators. During the training of the surrogate model, the input is the process parameter vector x, and the output is the quality indicator y. The training minimizes the prediction error, using 80% of the historical data for training and 20% for validation. The trained surrogate model has fast inference capabilities, with a single prediction taking milliseconds, meeting the simulation real-time requirements. By combining the mechanistic model and the surrogate model, the mechanistic model is used to ensure physical consistency for parts that can be explained by mechanistics, while the surrogate model is used to ensure accuracy for complex nonlinear parts, thus constructing a hybrid simulation model.

[0106] The digital twin model is dynamically calibrated and updated based on real-time data streams. When new production batch data is collected, the deviation between the model's predicted values ​​and actual measurements is calculated. If the deviation exceeds a threshold, model parameter updates are triggered. Bayesian inference or Kalman filtering methods are used to update the posterior distribution of model parameters based on the new data, achieving online adaptive modeling. The digital twin model operates synchronously with the actual industrial chain, forming a virtual-real mapping relationship and providing a high-fidelity environment for simulation verification of optimization schemes.

[0107] Step 5.2 establishes a hierarchical decision-making framework for collaborative optimization of supply chain quality. Based on the digital twin model constructed in Step 5.1 and the root cause diagnosis results output in Step 4, a hierarchical decision-making framework for collaborative optimization of supply chain quality is established. This framework decomposes the complex multi-enterprise, multi-parameter collaborative optimization problem into two levels: an upper-level global coordination layer and a lower-level local execution layer, reducing the difficulty of solving the problem. The upper level is the global coordination layer, operated by the quality management center of the supply chain alliance or a third-party coordination platform. It is responsible for formulating targeted overall supply chain quality objectives based on the root causes identified in Step 4, allocating quality sub-objectives for each link, and coordinating resource allocation and benefit distribution. The lower level is the local execution layer, operated by the intelligent agents of each enterprise or process link. It is responsible for optimizing local process parameters and operational decisions based on the quality objectives and resource constraints allocated by the upper level.

[0108] The inputs to the upper-level global coordination layer are the root cause diagnosis ranking list and recommended solutions output from step 4, the overall quality target of the industry chain, cost budget constraints, delivery constraints, and the critical path of quality influencing factors output from step 3. The optimization objective of the global coordination layer is: under the premise of meeting cost and delivery constraints, prioritize the development of improvement measures for the root cause nodes identified in step 4, and maximize the achievement of the overall quality target of the industry chain and effectively eliminate the identified root causes of quality anomalies by rationally allocating quality improvement tasks to each link. A multi-objective optimization model is established, with the objective function including a quality achievement term and a cost consumption term. The global objective function is calculated by multiplying the quality weight by the quality deviation term, and adding the cost weight multiplied by the cost term. The quality deviation term is the square of the difference between the optimized predicted value (summed over all terminal quality indicators) and the target value; the cost term is the improvement cost (summed over all links). The constraints include: transitive constraints on quality sub-objectives at each stage, i.e., downstream quality objectives impose requirements on upstream based on causal relationships in the quality graph; physical feasible domain constraints on process parameters at each stage, i.e., parameter values ​​must be within the equipment's capacity; total cost constraints, i.e., the sum of improvement costs across all stages must not exceed the budget; and delivery time constraints, i.e., the implementation time of the improvement plan must not exceed the deadline. To avoid deviations caused by directly adding different dimensions, normalized objective forms are used for quality deviation, cost, and the global objective function.

[0109] Based on the root cause diagnosis results output in Step 4, improvement measures are prioritized for the top 3 to 5 root cause nodes with the highest comprehensive scores. Combining the quality influencing factors identified in Step 3 and their degree of influence, the quality improvement potential and cost for each stage are determined. Root cause nodes identified in Step 4 are assigned higher improvement priority weights. For key influencing factors, their improvement potential is defined as the improvement amount of a unit change in the factor on the final quality target, i.e., the sensitivity coefficient calculated in Step 3.5. The improvement cost is defined as the additional cost required to control the factor within a more stringent range; the cost has a non-linear relationship with the improvement in control accuracy, using a marginal cost increasing model. The step to calculate the improvement cost is: add the basic cost to the marginal cost coefficient multiplied by the square of the improvement in control accuracy. Based on the improvement potential, cost, and root cause diagnosis priority, the comprehensive cost-benefit ratio of each factor is calculated, with root cause nodes and factors with high cost-benefit ratios prioritized for improvement tasks.

[0110] An improved Multi-Objective Particle Swarm Optimization (MOPSO) algorithm is used to solve the global optimization model. The particle swarm is initialized, with each particle representing an optimization scheme. The particle's position vector contains the target control range for each key factor in each stage, such as controlling the Dy content of rare earth ore at 0.8±0.05% and the sintering temperature at 1080±3℃. In each iteration, the achievement of the quality target and cost consumption of each particle are evaluated based on a digital twin simulation model, and the fitness function value is calculated. The individual optimal position of each particle and the set of globally optimal positions of the swarm are updated; that is, MOPSO maintains the Pareto optimal solution set. The velocity and position of the particles are updated based on the individual optimal and global optimal values. The position update step is: add the updated velocity to the current position to obtain the particle position for the next iteration. The iterations are repeated 500 to 2000 times until the Pareto front converges, yielding a set of non-dominated solutions, i.e., equilibrium solutions where both quality and cost objectives cannot be improved simultaneously. From the Pareto solution set, the final global optimization scheme is selected according to the decision-maker's preferences. The global optimization scheme includes the allocation of quality sub-objectives and resource quotas for each stage.

[0111] Step 5.3, Process agent decision-making based on local optimization and reinforcement learning. Based on the quality sub-objectives and resource quotas allocated by the global coordination layer in Step 5.2, the local agents in each process stage perform local process parameter optimization and operational decisions.

[0112] A local optimization problem is established for each stage, with the optimization objective being to minimize the deviation between the actual quality index and the sub-objective, plus the cost of process adjustments. The calculation steps for the local objective function are as follows: sum the quality indices of all sub-objectives, with the contribution of each index being its weight multiplied by the square of the difference between the actual and target values, plus the adjustment cost. The adjustment cost is related to the degree to which the process parameters deviate from the current set values; the larger the parameter adjustment range and the higher the frequency, the greater the cost. Constraints include: the physical feasible region of the process parameters; energy consumption constraints, with total energy consumption not exceeding the energy quota; and production cycle time constraints. To ensure the consistency of dimensions for different indices and costs, the local objectives can be normalized, including quality deviation normalization, adjustment cost normalization, and local objective function normalization.

[0113] Bayesian optimization is employed to solve local optimization problems. Bayesian optimization is suitable for costly black-box function optimization scenarios, where evaluating the effect of each combination of process parameters requires actual production or digital twin simulation, resulting in high costs. Bayesian optimization constructs a probabilistic surrogate model of the objective function, using information from existing evaluation points to guide the selection of the next evaluation point, finding the optimal solution with fewer evaluation iterations. During initialization, Latin hypercube sampling is performed in the process parameter space to select several initial evaluation points, such as 10 to 20 points. For each point, its quality indicators and costs are evaluated through digital twin simulation or small-batch trial production, obtaining an initial dataset. A Gaussian process model is trained based on this initial dataset. The Gaussian process (GP) predicts the mean and variance of the objective function value for any point in the parameter space. In each iteration, an EI (Extended Identification) function is used to select the next evaluation point in the parameter space. The EI function measures the expected improvement at a given point compared to the current optimal value. The steps for calculating the expected improvement are as follows: subtract the predicted mean from the current optimal target value, multiply the result by the standard normal cumulative distribution function applied to the standardized improvement magnitude, and add the predicted standard deviation multiplied by the standard normal probability density function applied to the standardized improvement magnitude. Select the point with the largest EI value as the next evaluation point, perform simulation or experimental evaluation to obtain the actual target value, add the evaluation point and target value to the dataset, and update the GP model. Repeat the iteration for 50 to 100 rounds. Since only one evaluation is required per round, the total number of evaluations is controllable, and finally, a locally optimal combination of process parameters is obtained.

[0114] Step 5.4, Iterative Coordination and Digital Twin Simulation Verification. Based on the global coordination scheme in Step 5.2 and the local optimization schemes for each stage in Step 5.3, iterative coordination and digital twin simulation verification are performed between the upper and lower layers. The process parameter schemes obtained from the local optimization of each stage are summarized and reported to the global coordination layer. The global coordination layer performs simulation verification on the overall scheme based on the digital twin model to evaluate whether the combination of schemes for each stage can achieve the overall quality target of the industrial chain. The simulation verification process is as follows: the optimized process parameters of each stage are input into the digital twin model to simulate the entire process from raw material input to product output and predict the distribution of quality indicators of the end product. If the simulation results show that the quality target achievement meets the requirements and the total cost is within the budget, the optimization scheme passes the verification and enters the implementation stage; if the simulation results are not ideal, the bottleneck links are analyzed, the allocation of quality sub-targets or resource quotas is adjusted at the global coordination layer, and then reissued to each stage for local optimization, forming an iterative cycle between the upper and lower layers. The iterative coordination process employs distributed optimization algorithms such as dual decomposition or the Alternating Direction Method of Multipliers (ADMM) to ensure that the iteration converges to the globally optimal or near-optimal solution.

[0115] This step outputs a collaborative optimization plan for supply chain quality, including the ID of each link or agent, the process parameter adjustment plan for that link, the allocated quality sub-objectives, the allocated resource quotas, and specific improvement measures for root cause nodes. The process parameter adjustment plan includes the target set values ​​and adjustment ranges for each parameter. A digital twin simulation verification report is also provided, including the predicted achievement rate of quality objectives, the estimated effect of root cause elimination, the cost estimate, and the robustness analysis results. Finally, the implementation schedule and responsibility allocation for the optimization plan are also provided.

[0116] Step 6: Implement the supply chain quality collaborative optimization plan in stages and conduct continuous monitoring and feedback to obtain an optimization implementation effect report;

[0117] Step 6.1: Phased Implementation and Process Monitoring of the Optimization Plan. Based on the supply chain quality collaborative optimization plan output in Step 5, a phased implementation plan is formulated and the process is monitored. The implementation process is divided into several phases according to the priority, interdependencies, and resource constraints of the improvement tasks at each stage. The first phase prioritizes emergency improvement measures targeting the root cause nodes identified in Step 4, as well as improvement tasks with high input-output ratios, low risk, and quick results, such as adjusting the quality acceptance standards of upstream raw materials and optimizing the control precision of key process parameters. The second phase implements medium-term tasks requiring equipment upgrades or personnel training. The third phase implements long-term tasks requiring significant investment or involving multi-party coordination. A detailed implementation Gantt chart is developed, clearly defining the start time, duration, responsible person, required resources, and quality acceptance standards for each task.

[0118] The process parameter adjustment instructions for the optimized solution are distributed to the Manufacturing Execution System (MES) or Distributed Control System (DCS) at each stage through the information system. For highly automated stages, the parameter adjustment instructions are automatically written into the control system for execution; for stages requiring manual operation, operation instructions are pushed to operators via mobile terminals or workstations. During the implementation of the solution, real-time quality data, process data, and equipment data are continuously collected to update the status information of the corresponding nodes in the metal industry chain quality map constructed in step 2. The collected real-time data includes: actual executed values ​​of process parameters at each stage, quality inspection results of intermediate and final products, equipment operating status and alarm information, production cycle time and output information, etc.

[0119] Establish a real-time quality monitoring dashboard to visually display the quality status and optimization progress of each link in the supply chain. The dashboard includes multiple views: a quality indicator trend chart, showing the curves of quality indicators changing over time, comparing data before and after optimization, and marking the target control range; a link status map, displaying the current status of each link in the form of a supply chain flowchart, with status types including normal, warning, and abnormal; a parameter control chart, displaying the SPC control chart of key process parameters, monitoring whether the parameters remain within the optimized target range; and an abnormal warning list, displaying detected quality abnormalities and equipment alarms in real time, sorted by severity. The monitoring dashboard supports drill-down functionality; clicking on a link or indicator allows viewing detailed information and historical data. Quality management personnel and process engineers can use the monitoring dashboard to keep abreast of the implementation of optimization plans and promptly identify and handle abnormalities.

[0120] Step 6.2, Optimization Effect Evaluation and Quality Improvement Calculation. Based on the implementation process data collected in Step 6.1, the actual effect of the optimization plan is quantitatively evaluated. A period before the implementation of the optimization plan is selected as the baseline period. Quality indicator data within the baseline period are collected as a comparison benchmark, and statistical quantities such as the mean, standard deviation, pass rate, and process capability index of each quality indicator are calculated. A period after the implementation of the optimization plan is selected as the evaluation period, such as 3 months, ensuring sufficient data sample size and a stable production process. Quality indicator data within the evaluation period are collected, and the statistical quantities for the evaluation period are calculated.

[0121] Compare quality indicators before and after implementation to calculate the degree of quality improvement. For quality indicators whose goal is to reduce fluctuations, the degree of improvement is measured by the percentage decrease in standard deviation. For indicators whose goal is to improve the pass rate, the degree of improvement is the absolute percentage increase in the pass rate. For indicators whose goal is to bring the mean closer to the target value, calculate the absolute value of the difference between the mean and the target value before implementation, then calculate the absolute value of the difference between the mean and the target value after implementation. Subtract the latter from the former and divide by the former; a positive value indicates that the mean is closer to the target. For process capability indices, subtract the process capability index before implementation from the process capability index after implementation, and divide by the process capability index before implementation. An increase in the optimized process capability index indicates improved process capability.

[0122] A comprehensive evaluation of the overall quality level of the industrial chain is conducted, establishing a Quality Score Index (QSI). This index considers multiple quality dimensions. Standardized scores are calculated based on the indicator's pass rate, mean deviation, and volatility, using min-max standardization to ensure scores range from 0 to 1, with 1 indicating the indicator has reached its optimal level. For benefit-type indicators, the original measurement is subtracted from the minimum value of the reference interval, and then divided by the difference between the maximum and minimum values ​​of the reference interval. For cost-type or deviation-type indicators, the maximum value of the reference interval is subtracted from the original measurement, and then divided by the difference between the maximum and minimum values ​​of the reference interval. The QSI values ​​before and after optimization are calculated, and the improvement in the QSI value reflects the overall quality improvement effect of the industrial chain. In addition to quality indicators, the impact of the optimization scheme on other dimensions such as cost, efficiency, and energy consumption is also assessed, conducting a multi-dimensional comprehensive benefit evaluation.

[0123] The optimization effect is statistically significant to determine whether the quality improvement is statistically significant rather than random fluctuation. A two-sample t-test or Mann-Whitney U-test is used to compare the distribution of quality indicators before and after optimization to see if there are significant differences. The null hypothesis is that the means of quality indicators are the same before and after optimization, and the alternative hypothesis is that the means of quality indicators after optimization are better than before. The t-statistic or U-statistic and its corresponding p-value are calculated. If the p-value is less than the significance level α, the null hypothesis is rejected, and the optimization effect is considered good. Statistical tests ensure the reliability of the evaluation conclusions.

[0124] Generate an optimization implementation effect report, which includes: optimization goal achievement, listing each quality goal and its actual achievement; quality improvement metrics, showing the improvement rate and statistical significance test results for each quality indicator; cost-benefit analysis, comparing the actual cost input with the economic benefits brought by quality improvement, such as reducing scrap loss and improving customer satisfaction; a summary of successful experiences of the optimization plan, extracting best practices that can be replicated and promoted; and existing problems and improvement suggestions, pointing out the problems found in the optimization process and future improvement directions.

[0125] Step 6.3: Online Model Updates and Knowledge Accumulation Based on Implementation Results. Based on the optimization implementation results report from Step 6.2, the models from Steps 3 and 5 are updated online to enable continuous learning and improvement of the system. New data collected during the optimization implementation period (including adjusted process parameters, actual quality results, and abnormal events) are added to the training dataset to retrain or incrementally update the graph neural network quality correlation analysis model. Incremental updates employ online learning algorithms, incorporating new knowledge while retaining existing knowledge to avoid catastrophic forgetting. For reinforcement learning models, new state-action-reward-next state quadruplets are added to the experience replay buffer for online policy updates, enabling the model to adapt to process changes and environmental drift. The reward function weights of the model are updated, and the trade-off coefficients between multiple objectives such as quality goals, cost constraints, and efficiency requirements are adjusted according to the actual results, making the model's decision preferences more consistent with actual management objectives.

[0126] The digital twin simulation model is calibrated and corrected. Simulation predictions are compared with actual measurements to calculate the prediction error. If the error exceeds an acceptable range, the model parameters are updated. For empirical coefficients in the mechanistic model, parameter identification methods are used to refit the coefficients based on new data to improve model accuracy. For data-driven surrogate models, new samples are added for retraining, and neural network weights or Gaussian process hyperparameters are updated. Through continuous calibration, the digital twin model is ensured to remain synchronized with the actual industry chain, providing an accurate simulation environment for subsequent optimization.

[0127] The successful experiences and lessons learned from the optimization schemes are distilled into knowledge, updating the metal industry chain quality map constructed in step 2. For successful optimization cases, new knowledge nodes and relationships are created in the quality map: nodes of the "Quality Improvement Case" type are created, with attributes including case description, application scenario, improvement measures, and effect data; relationship edges are established between case nodes and related problem nodes, cause nodes, measure nodes, and effect nodes to form a complete case knowledge structure; weights and confidence levels are labeled for relationship edges, determined based on the significance of actual effects. For new quality influencing factors or new causal relationships discovered during the optimization process, corresponding entity nodes and relationship edges are added to the map to enrich the quality knowledge base. For verified effective quality control rules, such as "when the Dy content exceeds 0.3wt%, intrinsic coercivity increases but costs increase," they are stored in the map as rule nodes to support subsequent reasoning applications.

[0128] This step outputs an optimization implementation effect report, including quality improvement metrics, goal achievement, cost-benefit analysis, and experience summary; an updated quality map with added successful case knowledge, newly discovered causal relationships, and verified effective control rules; an updated quality correlation analysis model, reinforcement learning decision model, and digital twin simulation model; a real-time quality monitoring dashboard and early warning system; and a continuous improvement plan and goals for the next round of optimization.

[0129] A computer-readable storage medium for storing computer-readable instructions that, when read by a computer, can execute a method for collaboratively improving the quality of a key metal industry chain based on a quality map, as described above.

[0130] In one embodiment of the present invention, a specific example is provided:

[0131] A rare earth permanent magnet material industry chain alliance includes a rare earth mining company (responsible for rare earth ore mining), two rare earth smelting companies (A and B, both responsible for rare earth separation and purification), five permanent magnet material manufacturing companies (responsible for NdFeB permanent magnet material production), and three motor manufacturing companies (responsible for manufacturing drive motors for new energy vehicles). The main quality problem facing this industry chain in 2025 is the large batch-to-batch fluctuation in the intrinsic coercivity of the finished NdFeB permanent magnet materials. The normal target is 1100 kA / m, but the actual fluctuation range reaches 1050 to 1150 kA / m, or ±50 kA / m, leading to a risk of demagnetization in some batches under the high-temperature operating conditions (150℃) of motors. Downstream motor companies require the intrinsic coercivity control range to be narrowed to within 1100 ± 30 kA / m, and the pass rate to be increased from the current 92% to over 98%.

[0132] Applying step 1 of this method, data acquisition systems were deployed at each stage of the industrial chain. Key component data of rare earth ore and separated products are shown in Table 1:

[0133] Table 1: Example of key component data for rare earth raw ore and separated products;

[0134]

[0135] Table 2 shows examples of sintering processes and magnetic property data for permanent magnet materials.

[0136] Table 2: Examples of sintering process and magnetic property data for permanent magnet materials;

[0137]

[0138] The collected data shows that the Dy content of rare earth ore fluctuates from 0.71% to 0.85%, and the sintering temperature fluctuates from 1075 to 1085℃. These fluctuations in upstream factors lead to a large fluctuation in the terminal Hcj from 1055 to 1118 kA / m.

[0139] The solution was implemented in January 2025. After a three-month transition period and stable operation, the results were satisfactory by April 2025. After optimization, the mean intrinsic coercivity was 1102 kA / m, the standard deviation decreased to 16 kA / m, the control range reached 1100 ± 28 kA / m, and the pass rate increased to 98.5%, exceeding the expected target. The improved quality stability reduced the demagnetization risk batches for motor manufacturers from 8% to below 1.5%, enhancing customer satisfaction and product competitiveness. This method achieved the goal of cross-enterprise collaborative quality improvement.

[0140] The embodiments of the present invention have been described above. However, the embodiments are not limited to the specific implementation methods described above. The specific implementation methods described above are merely illustrative and not restrictive. Those skilled in the art can make more equivalent embodiments under the guidance of the present embodiments, and all of them are within the protection scope of the present embodiments.

Claims

1. A method for collaboratively improving the quality of a key metal industry chain based on quality maps, characterized in that, include: Collect quality data from the metal industry chain, perform preprocessing, and obtain a standardized dataset; Based on the quality knowledge extracted from the standardized dataset, a preliminary quality map was constructed, and the quality map was fused and completed to obtain a quality map of the metal industry chain. Based on the quality map of the metal industry chain, we explore the correlation of quality influencing factors and obtain a set of quality influencing factors. Based on a set of quality influencing factors, intelligent diagnosis of quality anomalies is obtained by integrating statistical detection and graph inference to obtain root cause diagnosis results. Based on the root cause diagnosis results, the quality map of the key metal industry chain, and the set of key quality influencing factors, a digital twin simulation model is constructed. A hierarchical decision-making framework is used to coordinate and optimize each link of the industry chain, resulting in a collaborative optimization scheme for the quality of the industry chain. The supply chain quality collaborative optimization plan is implemented in stages and continuously monitored and feedback is conducted to obtain an optimization implementation effect report.

2. The method for collaborative quality improvement of a key metal industry chain based on quality maps according to claim 1, characterized in that, The preprocessing includes: Edge data acquisition nodes are deployed locally in each enterprise, and quality-related data are collected in real time using industrial communication protocols to obtain the raw data set of each enterprise. Data integrity checks were performed on the original datasets of each enterprise to identify missing values, duplicate records, and format errors. Potential outliers were marked using a box plot-based outlier detection method. Standardize the data format to obtain a local standardized dataset.

3. The method for collaborative quality improvement of a key metal industry chain based on quality maps according to claim 1, characterized in that, The preliminary construction of the quality map based on quality knowledge extracted from standardized datasets includes: Collect specialized corpora in the metals field to construct a domain corpus; Preprocess the domain corpus; Set up the basic framework of the domain pre-trained language model, perform self-supervised pre-training on the domain corpus, and adopt the masked language model MLM task; The pre-training process is carried out on a multi-GPU server. The number of training rounds is set, and the training progress is monitored based on the perplexity of the validation set. Training is stopped when the perplexity decreases. The pre-trained language model for the metal domain is obtained. Based on a pre-trained language model in the metal domain, we perform tasks such as quality entity recognition and relation extraction.

4. The method for collaborative quality improvement of a key metal industry chain based on quality maps according to claim 1, characterized in that, The factors influencing the quality of excavation include: Based on the quality map of the metal industry chain, a training task set of the meta-learning framework is constructed. The problem of quality influencing factor correlation mining is defined as a given target quality attribute node, the upstream factor node set is identified from the map, and the influence degree of each factor is quantified. A meta-model for quality correlation mining is trained using a model-independent meta-learning algorithm, and the meta-model employs a heterogeneous graph attention network. Heterogeneous graph neural network modeling is performed on the quality map of the metal industry chain. The embedded representation of the nodes is learned, and the quality map of the metal industry chain is modeled as a heterogeneous graph. The set of node types and the set of relationship types are defined.

5. A method for collaboratively improving the quality of a key metal industry chain based on quality maps, as described in claim 1, is characterized in that... The intelligent diagnosis of quality anomalies, based on a set of quality influencing factors and integrating statistical detection and graph inference, yields the following root cause diagnosis results: Based on real-time quality data streams and combined with a set of quality influencing factors, the statistical process control (SPC) method is used to detect anomalies in real-time quality data. The 3σ criterion is used to determine control limits, and when real-time quality data points exceed the control limits, they are marked as anomalies. Based on the detected set of abnormal quality indicators, combined with the metal industry chain quality map and the set of quality influencing factors, the graph reasoning method is used to trace the root cause of the abnormality. Starting from the abnormal node, the upstream cause node is identified by tracing back along the causal relationship edge, the process relationship edge, and the composition relationship edge. Based on the candidate cause nodes obtained through reverse tracing and the characteristics of the current anomaly, similar cases are retrieved from the historical anomaly case library to assist in root cause diagnosis, and the similarity between the current anomaly subgraph and the historical case subgraph is calculated.

6. The method for collaborative quality improvement of a key metal industry chain based on quality maps according to claim 1, characterized in that, The simulation of the synergistic effect of the optimized process parameter combinations in each stage on the entire industrial chain includes: Based on the root cause diagnosis results and recommended solutions, and combined with the metal industry chain quality map and the set of quality influencing factors, a digital twin simulation model of the metal industry chain is constructed. Process mechanism knowledge is extracted from the metal industry chain quality map, mechanism simulation models of each link are constructed, and the mechanism models of each link are connected and combined to construct an end-to-end simulation process of the industry chain. Establish a hierarchical decision-making framework for collaborative optimization of supply chain quality, decomposing the multi-enterprise, multi-parameter collaborative optimization problem into upper-level global coordination and lower-level local execution; Iterative coordination and digital twin simulation verification are carried out between upper and lower layers. The global coordination layer performs simulation verification of the overall scheme based on the digital twin model, inputs the optimized process parameters of each link into the digital twin model, simulates the entire process from raw material input to product output, and predicts the distribution of quality indicators of the end product.

7. A method for collaboratively improving the quality of a key metal industry chain based on quality maps, as described in claim 1, is characterized in that... The phased implementation of the supply chain quality collaborative optimization plan and the continuous monitoring and feedback include: Based on the supply chain quality collaborative optimization scheme, a phased implementation plan was formulated and the process was monitored. The implementation process was divided into several phases according to the priority, interdependence and resource constraints of the improvement tasks in each link. Based on the data collected during the implementation process, the actual effect of the supply chain quality collaborative optimization scheme is quantitatively evaluated. A baseline period and an evaluation period are set, and the quality indicators before and after implementation are compared to calculate the degree of quality improvement. Based on the optimization implementation effect report, the quality correlation analysis model is updated online. New data collected during the optimization implementation period is added to the training dataset, and the graph neural network quality correlation analysis model is retrained or incrementally updated. The successful experiences and lessons learned from the optimization scheme are precipitated into knowledge, and the quality map of the metal industry chain is updated.

8. A method for collaboratively improving the quality of a key metal industry chain based on quality maps, as described in claim 1, is characterized in that... The process of collecting and preprocessing local quality data from various enterprises in the metal industry chain also includes: This system enables cross-enterprise collaborative data analysis based on a federated learning framework. Global data analysis model parameters are initialized on a central cloud server, and each enterprise trains a local model copy based on a local standardized dataset. Differential privacy processing is performed on the local gradients. After collecting the gradient information uploaded by all enterprises, the central server performs weighted aggregation to update the global model parameters.

9. A method for collaboratively improving the quality of a key metal industry chain based on quality maps, as described in claim 1, is characterized in that... The correlation of factors affecting mining quality also includes: To identify and quantify the impact of quality-influencing factors, a trained heterogeneous graph neural network model is used to identify key factor nodes and quantify the impact of each factor. An attention-weight-based method for identifying influencing factors is adopted. For a target quality attribute node, starting from the node, the process traces back along the edges of the graph to the upstream set of candidate influencing factor nodes. For each candidate factor node, the attention path weight of its relationship to the target node is calculated. Using sensitivity analysis, for the identified key factor nodes, while keeping the characteristics of other factor nodes unchanged in the test samples, only the feature value of the node is changed. The change in the target quality attribute value is predicted by a graph neural network model, and the sensitivity coefficient is calculated. By combining attention weight and sensitivity coefficient, a comprehensive influence score is defined. Based on the comprehensive score, key factors are ranked to obtain a set of quality influencing factors.

10. A computer-readable storage medium, characterized in that, It is used to store computer-readable instructions, which, when read by a computer, can execute the method for collaborative quality improvement of a key metal industry chain based on quality maps as described in any one of claims 1-9.

Images