Digital-based multi-mode production operation assistance and quality control method

Abstract

The application relates to the technical field of intelligent manufacturing and quality management, and provides a multi-mode production operation assisting and quality management method based on digitization, which comprises the following steps: acquiring multi-source heterogeneous data; for each pair of data sources, calculating mutual information within a preset delay parameter range, selecting a delay parameter corresponding to a maximum value as an optimal delay parameter for time shift compensation; constructing a quality knowledge graph with production elements as nodes and process flow relationship and quality causal correlation as edges, and determining an edge weight of an edge representing the quality causal correlation according to the mutual information; dynamically inferring and outputting a quality risk score through a graph neural network, generating a risk transmission path and a process adjustment strategy; and pushing the strategy to an operation terminal and returning an execution result as a feedback signal to update the graph neural network parameters. The application improves data alignment accuracy, realizes dynamic structured representation of quality knowledge, and realizes closed-loop adaptive optimization of reasoning and execution.

Description

Technical Field

[0001] This application relates to the field of intelligent manufacturing and quality management technology, specifically to a digitally-based multi-mode production operation assistance and quality control method. Background Technology

[0002] In multi-mode production scenarios, the same production line needs to flexibly switch product models and process routes, resulting in varying communication latency for heterogeneous data from multiple sources, such as equipment sensors, process control systems, and quality inspection devices. Existing technologies typically use fixed time windows or simple timestamp-based interpolation methods for data alignment. These methods rely on strict clock synchronization between data sources, making it difficult to adapt to dynamic latency changes caused by process route switching. This leads to insufficient data alignment accuracy, which in turn affects the reliability of subsequent quality analysis.

[0003] Furthermore, existing quality control methods largely rely on statistical process control or single machine learning models, whose internal decision-making processes are often opaque, lacking structured knowledge representation and interpretable risk transmission paths. While some solutions attempt to introduce knowledge graphs for anomaly tracing, these graphs primarily rely on offline batch processing for construction, resulting in updates lagging behind the real-time production line status. Moreover, they fail to address the time alignment issue of node data sources, leading to discrepancies between graph attributes and actual production conditions.

[0004] More importantly, existing quality inference models typically run with fixed parameters after training, lacking a closed-loop feedback mechanism between the model and the actual operations of on-site personnel. Quality control results cannot be iteratively optimized online based on operator performance, making it difficult for the model to adapt to changes in specific production line conditions. This results in a disconnect between operational support and quality control, limiting overall control efficiency. Therefore, there is an urgent need for a quality control method capable of accurately aligning multi-source data, constructing a dynamic and interpretable quality knowledge graph, and achieving closed-loop feedback between inference and execution. Summary of the Invention

[0005] To address the problems of low time alignment accuracy of multi-source data, static lag in quality knowledge representation, and lack of closed-loop feedback in reasoning and execution in existing technologies, this application proposes a digitally based multi-mode production operation assistance and quality control method to achieve high-precision adaptive data alignment, dynamic and interpretable quality risk reasoning, and online adaptive closed-loop control.

[0006] To achieve the above objectives, this application adopts the following technical solution:

[0007] A digitally-based multi-modal production operation assistance and quality control method includes the following steps:

[0008] Acquire multi-source heterogeneous data from a multi-mode production process. The multi-source heterogeneous data includes equipment operating parameters, process data, environmental condition data, and quality inspection data. For each pair of data sources, calculate the mutual information between two data sequences under different delay parameters within a preset delay parameter range. Select the delay parameter that reaches the maximum value of the mutual information as the optimal delay parameter between the pair of data sources. Perform time shift compensation on the two data sequences based on the optimal delay parameter to obtain time-aligned multi-source heterogeneous data.

[0009] Using production factors as nodes and process flow relationships and quality causal relationships as edges, a quality knowledge graph is constructed based on the time-aligned multi-source heterogeneous data. The production factors include equipment, processes, materials, and quality indicators. The edges include edges representing process flow relationships and edges representing quality causal relationships. The edge weights of the edges representing quality causal relationships are determined based on the mutual information of the connected nodes in the time-aligned data sequence.

[0010] The graph neural network performs dynamic reasoning based on all nodes and edges in the quality knowledge graph, outputs a quality risk score for each node, and generates a quality risk transmission path and process adjustment strategy based on nodes whose quality risk scores exceed a preset threshold.

[0011] The quality risk transmission path and the process adjustment strategy are pushed to the operation terminal. The execution result of the process adjustment strategy by the operator is sent back as a feedback signal. The parameters of the graph neural network are updated online based on the feedback signal.

[0012] As one implementation, the feature vector of the node in the quality knowledge graph contains the statistical features of the time-aligned data sequence, including mean, variance, trend slope, and peak interval; the edges of the quality knowledge graph are divided into explicit edges and implicit edges, the explicit edges represent known process flow relationships, and the implicit edges represent the association relationships between node pairs whose mutual information exceeds a preset mutual information threshold.

[0013] As one implementation method, the mutual information is calculated as follows: for any two data sequences X and Y, with a delay parameter Next, calculate the joint probability distribution. With marginal probability distribution and The relative entropy between them, the formula for calculating the relative entropy is:

[0014]

[0015] Where X and Y are random variables, and x and y are the specific values ​​that X and Y take, respectively. For delay parameters, The mutual information, expressed in bits, is in the delay parameter. The value below, For delay parameters The joint probability distribution when X takes the value x and Y takes the value y. Let X be the marginal probability distribution when X takes the value x. Let Y be the marginal probability distribution when Y takes the value y. This represents the logarithm to base 2, summing all possible values ​​of x and y, and selecting the value that makes the sum... Reaching the maximum value This serves as the optimal delay parameter.

[0016] As one implementation method, the process adjustment strategy is generated based on the adjustable process parameters of each node in the quality risk transmission path; the feedback signal is the change in quality indicators before and after the operator executes the process adjustment strategy.

[0017] As one implementation method, the quality risk transmission path is generated as follows: starting from the node where the quality risk score exceeds the preset threshold, the path with the largest edge weight in the quality knowledge graph is traversed in reverse until the material node or equipment node is reached, and the sequence of nodes traversed in the reverse traversal is taken as the quality risk transmission path.

[0018] As one implementation method, the method further includes: acquiring process document data from the multi-mode production process; performing semantic parsing on the process document data using a multimodal large language model to extract process routes, procedures, steps, and process parameter constraints; and injecting the extraction results into the quality knowledge graph in the form of triples as new nodes and edges; the multimodal large language model receives the text content of the process document and associated product image data, and performs confidence evaluation on the parsed triples, injecting only triples with confidence scores higher than a preset confidence threshold into the quality knowledge graph.

[0019] As one implementation, when the graph neural network outputs a node whose quality risk score exceeds the preset threshold, the associated data of that node and the neighbor subgraph of that node in the quality knowledge graph are input into the multimodal large language model. The multimodal large language model performs root cause analysis on the quality deviation based on parameterized knowledge, outputs a root cause description and recommended corrective measures in natural language form, and verifies the consistency between the root cause description and the corrective measures and the quality risk transmission path.

[0020] As one implementation method, the method further includes: constructing a federated learning framework, which includes a central server and local clients deployed at multiple production sites; each local client iteratively training a local quality prediction model based on multi-source heterogeneous data aligned to local time, wherein the input of the local quality prediction model is the feature vector of a node in the quality knowledge graph, and the output is the quality risk score; and each local client uploading the gradient parameters obtained from the iterative training to the central server after homomorphic encryption.

[0021] In one implementation, the central server securely aggregates the gradient parameters uploaded by each local client to obtain the parameters of the global quality prediction model, and then distributes the parameters of the global quality prediction model to each local client. Each local client uses the global quality prediction model as the teacher model and a locally deployed lightweight model as the student model. By minimizing the divergence loss between the output distribution of the teacher model and the output distribution of the student model, knowledge transfer from the global quality prediction model to the student model is achieved.

[0022] As one implementation, when a new production line or new product model is introduced, the local client performs rapid fine-tuning of the student model on a small number of local samples based on a model-independent meta-learning framework. The model-independent meta-learning framework enables the student model to quickly adapt to the new data distribution under a small number of samples by calculating task-specific gradients based on support set samples in the inner loop and updating meta-parameters based on query set samples in the outer loop.

[0023] The beneficial effects of this application are as follows: By adopting a time alignment algorithm based on the mutual information maximization criterion, this application adaptively determines the optimal delay parameter by utilizing the statistical correlation between data sequences, overcoming the limitations of traditional methods that rely on clock synchronization assumptions, and improving the data alignment accuracy under dynamic delays in multi-mode production; On this basis, the edge weights representing the quality causal relationships in the quality knowledge graph are determined using the same mutual information, so that the edge weights are dynamically determined according to the production line data status, and the graph representation is consistent with the actual production line status, solving the problems of existing graph update lag and node attribute deviation, and realizing the dynamic structured representation of quality knowledge; Furthermore, dynamic reasoning is performed through graph neural networks to generate traceable risk transmission paths, and the execution results of the operator's process adjustment strategy are used as feedback signals to update the parameters of the graph neural network, forming a closed-loop mechanism of reasoning and execution, realizing the online continuous adaptive optimization of the quality control model for specific production line conditions, effectively improving the accuracy of operation assistance and the overall efficiency of quality control. Attached Figure Description

[0024] Figure 1This is a flowchart of a method for digital multi-mode production operation assistance and quality control according to an embodiment of this application. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application.

[0026] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.

[0027] Embodiment 1 of this application is as follows:

[0028] like Figure 1 As shown, this embodiment provides a digital-based multi-mode production operation assistance and quality control method. This method aims to solve the problems of data asynchrony, static lag in knowledge representation, and the separation of reasoning and execution caused by process route switching in multi-mode production. It achieves online adaptive quality control by constructing a complete technical chain of "perception-reasoning-execution-feedback".

[0029] Step S100: Obtain multi-source heterogeneous data in the multi-mode production process. For each pair of data sources, calculate the mutual information between the two data sequences under different delay parameters within the preset delay parameter range. Select the delay parameter with the maximum mutual information as the optimal delay parameter between the pair of data sources. Perform time shift compensation on the two data sequences based on the optimal delay parameter to obtain time-aligned multi-source heterogeneous data.

[0030] Specifically, multi-source heterogeneous data includes equipment operating parameters, process data, environmental condition data, and quality inspection data. In multi-mode production scenarios, the sampling frequency, communication latency, and triggering mechanisms of different sensors and control systems vary, leading to inconsistent recording times of the same physical event in different data sources. Traditional fixed-time-window or timestamp-based interpolation methods rely on strict clock synchronization assumptions and are difficult to adapt to dynamically changing latency. This embodiment uses a mutual information maximization criterion based on statistical correlation to determine the alignment relationship. Mutual information reflects the degree of interdependence between two random variables. When two data sequences truly correspond to the same production event in time, their mutual information usually reaches its peak. Therefore, by searching for the delay parameter that maximizes mutual information within a preset range, asynchronous errors caused by network jitter or processing time can be adaptively eliminated. It should be understood that this only involves translation compensation in the time dimension and does not involve spatial coordinate transformation, hence the term "time alignment." For example, for temperature sensor data and vibration sensor data, if the mutual information is maximized when the vibration data lags behind the temperature data by 50ms, then the vibration data sequence is shifted forward by 50ms to align with the temperature data. This alignment method, based on statistical characteristics of the data, overcomes the accuracy loss caused by asynchronous hardware clocks, providing a high-confidence data foundation for subsequent analysis.

[0031] Step S200: Using production factors as nodes and process flow relationships and quality causal relationships as edges, construct a quality knowledge graph based on time-aligned multi-source heterogeneous data. The edges include edges representing process flow relationships and edges representing quality causal relationships. The edge weights of the edges representing quality causal relationships are determined based on the mutual information of the connected nodes in the time-aligned data sequence.

[0032] Specifically, production factors include equipment, processes, materials, and quality indicators. The structured representation of the quality knowledge graph makes complex production relationships traceable. Unlike traditional offline batch-constructed graphs with fixed weights, this embodiment uses the mutual information calculated in step S100 to directly determine the weights of edges representing causal relationships in quality within the graph. For explicit edges representing known process flow relationships, their weights can be set according to preset process constraints. This means that if two nodes (such as the temperature parameter of a process and the hardness index of the final product) exhibit a high statistical correlation (i.e., high mutual information) in the time-aligned data sequence, their edge weights are high, indicating a strong potential causal or correlational influence between them. Conversely, if the mutual information is low, the edge weights are low or even there is no implicit correlation. By reusing the same mutual information for both data alignment and graph weighting, logical coupling between the data layer and the knowledge layer is achieved. This enables the graph to be dynamically updated with the real-time data status of the production line, ensuring that the node attributes in the graph are consistent with the actual production status. This solves the problems of lagging updates and node attribute deviations in existing graphs, and realizes dynamic structured representation of quality knowledge.

[0033] Step S300: Dynamic reasoning is performed based on all nodes and edges in the quality knowledge graph using a graph neural network to output the quality risk score of each node. Quality risk transmission paths and process adjustment strategies are generated based on nodes whose quality risk scores exceed a preset threshold.

[0034] Specifically, a graph neural network (GNN) is a deep learning model specifically designed for processing graph-structured data. Its basic working principle involves aggregating feature information from neighboring nodes through a message-passing mechanism to update the representation of the central node. In this embodiment, the GNN receives a quality knowledge graph as input, where node features include time-aligned statistical features of the data sequence, and the edge weights representing causal relationships in quality are determined by mutual information. The network propagates through multiple layers, capturing dependencies within local subgraphs to predict the probability of quality anomalies at each node, i.e., a quality risk score. When a node's risk score exceeds a preset threshold, the system identifies it as a high-risk node. Subsequently, based on the graph topology and edge weights, it traces back to the source node, generating a quality risk transmission path that visually demonstrates how risk is transmitted from upstream processes or equipment to downstream quality indicators. Simultaneously, the system generates process adjustment strategies based on adjustable parameters along the path. For example, if the path shows that temperature fluctuations in a welding process lead to unacceptable final strength, the strategy might suggest fine-tuning the temperature setpoint for that process. This process transforms black-box prediction into interpretable path derivation, improving the credibility of the decision.

[0035] In step S400, the quality risk transmission path and process adjustment strategy are pushed to the operation terminal, and the execution results of the process adjustment strategy by the operators are sent back as feedback signals. The parameters of the graph neural network are updated online based on the feedback signals.

[0036] Specifically, the generated strategy is pushed to on-site operators via industrial tablets, AR glasses, or central control screens. After the operators implement the adjustments, the quality inspection system collects new quality data and quantifies the changes in quality indicators before and after the adjustments as feedback signals. These feedback signals are fed back to the graph neural network for online iterative optimization. Specifically, the feedback signals can be used as labels for supervised learning or reward functions for reinforcement learning, and the internal parameters (such as the weight matrix) of the graph neural network can be adjusted using algorithms such as gradient descent. This closed-loop mechanism allows the model to learn from actual production results and continuously correct its understanding of specific production line conditions. For example, if a strategy adjustment fails to effectively reduce risk, the model will lower the confidence of the relevant paths or adjust the attention distribution to avoid similar erroneous judgments in the next inference. This achieves an integrated closed loop of operational assistance and quality control, enabling the model to continuously adapt to dynamic production environments.

[0037] Embodiment 2 of this application is as follows:

[0038] Based on Example 1, this example further refines the time alignment process of multi-source heterogeneous data and the construction details of the quality knowledge graph.

[0039] Specifically, regarding the time alignment operation in step S100, the mutual information is calculated as follows: For any two data sequences X and Y, the mutual information is calculated using the delay parameter... Next, calculate the joint probability distribution. With marginal probability distribution and The relative entropy between them.

[0040] The formula for calculating relative entropy is:

[0041]

[0042] In this formula, X and Y are random variables, and x and y are the specific values ​​of X and Y, such as the reading of a temperature sensor or the amplitude of a vibration sensor. This is the delay parameter, representing the amount of time shift applied to one of the data sequences; Mutual information in bits in the delay parameter The value below is used to quantify the strength of the statistical correlation between two sequences at a specific time offset; For delay parameters The joint probability distribution when X takes the value x and Y takes the value y reflects the probability that the two variables take a specific value at the same time; Let X be the marginal probability distribution when X takes the value x. The marginal probability distributions for Y taking the value y, each reflecting the independent probability of a single variable taking a specific value; It represents a logarithm to the base 2; the summation symbol iterates through all possible values ​​of x and y.

[0043] In practical calculations, the above probability distribution can be approximated using histogram binning or kernel density estimation. The system iterates through multiple [variables] within a preset range. Values, calculate the corresponding values ​​respectively Select to make Reaching the maximum value As the optimal delay parameter.

[0044] This alignment mechanism, based on maximizing mutual information, operates on the principle that when two data sequences truly correspond to the same physical process, they possess the highest degree of interdependence—the peak mutual information—under correct time alignment. Compared to methods relying on fixed clock synchronization, this algorithm can adaptively compensate for asynchronous errors caused by network transmission jitter or differences in processing time, thereby obtaining highly accurate time-aligned data.

[0045] Specifically, for the graph construction operation in step S200, the feature vector of the node in the quality knowledge graph contains the statistical features of the time-aligned data sequence, including the mean, variance, trend slope, and peak interval.

[0046] Taking equipment operating parameters as an example, the mean reflects the average load level of the equipment over a period of time, the variance characterizes the stability of data fluctuations, the trend slope indicates the direction of parameter change over time (such as the rate of temperature rise), and the peak interval captures the frequency of periodic fault characteristics. These statistical features are extracted from the time-aligned raw data through a sliding time window and normalized to eliminate the influence of dimensions, ultimately forming the initial feature vector of the node.

[0047] Furthermore, the edges of a quality knowledge graph are divided into explicit edges and implicit edges.

[0048] Explicit edges represent known process flow relationships, such as the inevitable transition from process A to process B after completion. These edges are directly generated by the process route definition in the Manufacturing Execution System (MES), and their weights are usually set to a fixed high value or dynamically adjusted based on historical yield rates. Implicit edges represent the association between node pairs whose mutual information exceeds a preset mutual information threshold.

[0049] During the construction process, the system calculates the mutual information of the time-aligned data sequences between all potential node pairs. If the mutual information between two nodes (such as temperature sensor data from one process and quality inspection data from another process) exceeds a set threshold, an implicit edge is established between them, and the value of the mutual information is mapped to the edge weight.

[0050] In this way, the graph not only contains a defined process flow structure but also uncovers potential data-driven causal relationships. The edge weights representing quality causal relationships are dynamically updated with the real-time data status of the production line, enabling the graph to keenly capture abnormal coupling relationships in the production process. This provides rich input with both structured knowledge and data relevance for subsequent graph neural network inference.

[0051] Embodiment 3 of this application is as follows:

[0052] Based on Example 1, this example further refines the generation process of the quality risk transmission path in step S300 and the quantification mechanism of the process adjustment strategy and feedback signal in step S400.

[0053] Specifically, for the risk transmission path generation in step S300, starting from the node where the quality risk score exceeds the preset threshold, the process proceeds in reverse along the path with the largest edge weight in the quality knowledge graph until the material node or equipment node is reached. The sequence of nodes traversed in the reverse traversal is then used as the quality risk transmission path.

[0054] The execution logic of this process is as follows: The system first identifies a set of high-risk nodes with risk scores exceeding a threshold. For each high-risk node, a depth-first search or breadth-first search algorithm is initiated, but when selecting the next-hop neighbor node, a greedy strategy is strictly followed, prioritizing the neighbor node that connects to the current node and has the largest edge weight. Since the edge weights representing causal relationships in quality are determined by the mutual information of the time-aligned data sequence, high weights imply strong statistical correlation. Therefore, this traversal method can trace upstream along the causal chain that has the most significant impact on quality. The traversal process continues until a node of type "material" or "equipment" is encountered, as these two types of nodes typically represent the source or basic input of production factors and no longer possess upstream process attributes. The resulting node sequence, for example, tracing back from "finished product hardness failure" to "heat treatment furnace temperature fluctuation" and then to "abnormal thermocouple sensor readings," clearly demonstrates the physical transmission link of quality deviations. It should be understood that if multiple candidate paths with the same edge weight are encountered during the traversal, the system can perform a secondary sorting based on historical failure frequency or process importance coefficient, or output multiple potential paths in parallel for manual review, to ensure the comprehensiveness of the traceability.

[0055] Specifically, the process adjustment strategy in step S300 is generated based on the adjustable process parameters of each node in the quality risk transmission path.

[0056] Once the risk transmission path is determined, the system queries the process specification database associated with each node on the path and extracts the adjustable process parameters corresponding to that node. For example, if the path includes a welding process node, its adjustable parameters might include welding current, voltage, travel speed, and shielding gas flow rate; if it includes an injection molding process node, its adjustable parameters might include injection pressure, holding time, and mold temperature. Based on the inference results of the graph neural network, the system calculates the sensitivity contribution of each parameter to the final quality index, thereby generating specific adjustment suggestions. For example, if the analysis shows that the welding current has the greatest impact on strength and the current value deviates from the optimal range, the strategy suggests adjusting the current to the recommended range. This strategy generation method transforms abstract risk scores into parameter setting instructions that on-site operators can directly execute, realizing the implementation from data analysis to work guidance.

[0057] Specifically, the feedback signal quantification and model update in step S400 refer to the changes in quality indicators before and after the operator implements the process adjustment strategy.

[0058] In actual operation, after the operator completes the adjustment according to the pushed strategy, the quality inspection system collects new quality data from subsequent batches or the remaining products in the same batch. The system compares the average quality indicator (or specific defect rate) before the adjustment with the average quality indicator after the adjustment, calculates the difference or relative rate of change between the two, and defines this change as a feedback signal. For example, if the defect rate was 5% before the adjustment and decreased to 2% after the adjustment, the feedback signal represents a quality improvement of -3%.

[0059] The parameters of the graph neural network are updated online based on this feedback signal. Specific implementation methods include: First, online regression calibration: Before execution, the graph neural network outputs quality risk scores for each node, including risk scores for quality indicator nodes. After the operator executes the process adjustment strategy, the system normalizes the measured quality indicator data and constructs a regression loss based on the deviation between the predicted risk scores for quality indicator nodes before execution and the measured quality status after execution. The gradient of this loss with respect to the internal parameters of the graph neural network is calculated through backpropagation, and the parameters are updated accordingly, gradually aligning the model's prediction of the risk scores for quality indicator nodes with the actual execution results. Second, reinforcement learning reward method: The change in quality indicators is used as a reward signal. The decision-making process of generating and executing the process adjustment strategy is considered an action of the agent. The graph neural network is used as the policy evaluation network. The reward signal is used to estimate the reward for each node's score on the risk transmission path. The graph neural network parameters are updated through policy gradient or temporal difference methods, making the model biased towards outputting risk judgments and process adjustment strategies that bring positive quality improvement. In both of the above approaches, if the feedback signal indicates a significant improvement in quality, the gradient update direction will strengthen the mapping relationship between relevant edge weights and node features; conversely, if the quality does not improve or even deteriorates, the model will adjust the confidence in the relevant causal relationships and reduce the weight contribution of the corresponding paths. This process enables the graph neural network to continuously calibrate its understanding of multi-mode production conditions during online operation, achieving closed-loop adaptive optimization between the inference model and the on-site execution effect.

[0060] Embodiment 4 of this application is as follows:

[0061] Based on Examples 1 to 3, this example further describes how to integrate unstructured process knowledge into the main process flow, and the specific implementation method of enhancing interpretability by utilizing a multimodal large language model.

[0062] For the dynamic injection of process knowledge, process document data from multi-mode production processes is acquired. A multimodal large language model is used to perform semantic parsing on the process document data, extracting process routes, procedures, steps, and process parameter constraints. The extracted results are injected into the quality knowledge graph in the form of triples as new nodes and edges. The multimodal large language model receives the text content of the process documents and the associated product image data, and evaluates the confidence of the parsed triples. Only triples with confidence scores higher than a preset confidence threshold are injected into the quality knowledge graph.

[0063] Specifically, process documentation data typically includes unstructured text such as work instructions, equipment maintenance manuals, and product specifications, along with accompanying technical drawings or photographs of physical objects. The multimodal large language model processes image data through a visual encoder and text data through a text encoder, mapping both to a unified semantic space for joint reasoning. A "triple" refers to a structured knowledge unit consisting of a head entity, a relation type, and a tail entity, for example (head entity: "Injection Molding Machine A", relation: "Set Temperature Upper Limit", tail entity: "220℃"). The system calculates a confidence score for each triple generated by the model. This score can be based on the log-likelihood probability when the model generated the triple, or obtained through a consensus voting mechanism using multiple sampling. Only when the confidence score exceeds a preset threshold (e.g., 0.95) is the triple recognized as valid knowledge and written into the graph. This filtering mechanism effectively suppresses the "illusion" phenomenon of large models, preventing erroneous process parameters from being introduced into the decision-making process, thereby ensuring the accuracy and reliability of structured data in the quality knowledge graph.

[0064] For deep attribution of quality anomalies, when a node whose quality risk score output by the graph neural network exceeds a preset threshold is a node, its associated data and the neighboring subgraphs of that node in the quality knowledge graph are input into a multimodal large language model. The multimodal large language model performs root cause analysis on the quality deviation based on parameterized knowledge, outputs a root cause description and recommended corrective measures in natural language form, and verifies the consistency between the root cause description and corrective measures and the quality risk transmission path.

[0065] Specifically, when an abnormally high risk score is detected in a certain process or equipment node, the system not only relies on numerical prediction from a graph neural network but also activates a root cause analysis module. This module extracts a local subgraph consisting of the abnormal node and its directly connected neighboring nodes, and combines this with historical operating data, environmental parameters, and other related data to construct a prompt input to the multimodal large language model. The large language model uses its internally stored industry-wide knowledge and domain-specific knowledge learned from process documents to generate a natural language description explaining the possible causes of the anomaly (e.g., "insufficient cooling water flow leads to excessively high mold temperature") and providing suggested measures. Subsequently, a consistency verification step is performed: the system compares the key entities and operational logic mentioned in the natural language description generated by the large language model with the node sequence and edge weights in the "quality risk transmission path" generated by reverse traversal in Example 3. If the cause node mentioned by the large model is not on the risk transmission path, or if the causal direction described contradicts the correlation direction indicated by the edge weights in the graph, it is determined to be inconsistent, and the system will reduce the weight of the root cause analysis or trigger manual review. This verification mechanism achieves complementarity and balance between data-driven structured reasoning and knowledge-driven natural language reasoning. It leverages the generalization and explanatory power of large models while preserving the logical rigor of graph reasoning, significantly improving the credibility of quality control decisions.

[0066] Embodiment 5 of this application is as follows:

[0067] Based on Examples 1 to 4, this example further describes how to achieve collaborative optimization and rapid deployment of quality control models in a multi-factory or multi-production-line environment while protecting data privacy.

[0068] To address the problem of data silos from multiple sources, a federated learning framework is constructed, which includes a central server and local clients deployed in multiple production sites.

[0069] Specifically, each local client independently runs the aforementioned steps S100 to S400 to generate its own quality knowledge graph and node feature vectors. Each local client iteratively trains its local quality prediction model based on multi-source heterogeneous data aligned to local time. The input of the local quality prediction model is the feature vector of the node in the quality knowledge graph, and the output is a quality risk score.

[0070] This design ensures that raw, sensitive data (such as equipment parameters and process formulas) does not leave the factory area, and only model parameters are shared. To prevent the leakage of local data distribution information during parameter exchange, each local client uploads the gradient parameters obtained from iterative training to the central server after homomorphic encryption.

[0071] It should be understood that "iterative training" here refers to the periodic update behavior during online operation, rather than the pre-training in the offline stage, so as to ensure that the model can reflect the latest status of the production line in real time.

[0072] The central server securely aggregates the gradient parameters uploaded by each local client to obtain the parameters of the global quality prediction model, and then distributes the parameters of the global quality prediction model to each local client.

[0073] The secure aggregation process can employ federated averaging algorithms or weighted aggregation strategies to complete parameter fusion without decrypting gradients, thereby achieving knowledge sharing across plant areas while strictly protecting data privacy.

[0074] Subsequently, in order to adapt to the limited computing resources on the edge, each local client uses the received global quality prediction model as the teacher model and the locally deployed lightweight model as the student model. By minimizing the divergence loss between the output distribution of the teacher model and the output distribution of the student model, knowledge transfer from the global quality prediction model to the student model is achieved.

[0075] For example, Kullback-Leibler (KL) divergence can be used as a loss function to force lightweight student models to mimic the decision boundaries of teacher models. This process enables local terminals to achieve inference accuracy close to that of large global models under low computing power, solving the problem that complex graph neural networks are difficult to deploy directly on industrial field terminals.

[0076] When a new production line or new product model is introduced, training the model directly from scratch is inefficient and ineffective due to a lack of sufficient historical data. In this case, the local client uses a Model-Agnostic Meta-Learning (MAML) framework to quickly fine-tune the student model on a small number of local samples.

[0077] Specifically, the model-independent meta-learning framework enables student models to quickly adapt to new data distributions with a small number of samples by calculating task-specific gradients based on support set samples in the inner loop and updating meta-parameters based on query set samples in the outer loop.

[0078] In practice, the system collects a small amount of initial quality data from the new production line as a support set to simulate task gradients under different operating conditions; at the same time, it retains another set of data as a query set to evaluate the model's generalization ability and update meta-parameters. This mechanism gives the model the ability to "learn," enabling it to quickly converge to a high-precision state without extensive retraining when facing product models that frequently switch between multiple production modes, significantly improving the system's flexibility and response speed.

[0079] Embodiment 6 of this application is as follows:

[0080] To more clearly illustrate the technical solution of this application and its application effect in actual production, the following describes a complete process of quality anomaly handling and closed-loop optimization, taking the permanent magnet synchronous motor stator assembly line of a discrete manufacturing plant as an example.

[0081] In this scenario, the assembly line comprises multiple workstations, including winding machines, coiling machines, impregnation ovens, and a comprehensive testing platform. Each workstation is equipped with different types of sensors, including temperature sensors, vibration accelerometers, current transformers, and visual inspection cameras. Due to differing communication protocols and fluctuating network load, the winding tension data of the same batch of motors and the temperature data of the subsequent impregnation oven exhibit a random time delay of approximately 200ms to 500ms in the original records, rendering traditional timestamp-based alignment methods ineffective.

[0082] First, step S100 is executed to acquire data and align time. The system acquires the tension sequence X of the winding machine and the temperature sequence Y of the impregnation furnace in real time. For this pair of data sources, the system calculates the mutual information every 10ms within a preset delay parameter range [-1000ms, 1000ms]. In the formula, X represents the tension sequence of the winding machine, and Y represents the temperature sequence of the impregnation furnace. This is the delay parameter. The calculation results show that when... At this point, the mutual information reaches its peak, indicating that the statistical correlation between the two sequences is strongest, meaning that tension fluctuations are most likely to affect the subsequent coating temperature distribution. Based on this, the system performs time-shift compensation on the temperature sequences to obtain time-aligned data.

[0083] Next, step S200 is executed to construct the knowledge graph. The system maps production elements such as "winding process," "impregnation process," and "stator insulation resistance" into nodes. The node feature vector consists of the mean, variance, trend slope, and peak interval of the time-aligned data. For explicit edges, a "winding -> impregnation" connection is established according to the standard process route in the MES system; for implicit edges, the system detects that the mutual information between the "winding tension" node and the "impregnation bubble rate" node exceeds the preset threshold of 0.8, therefore, a high-weight implicit edge is established between the two to represent the potential causal relationship.

[0084] The system then performs dynamic reasoning and strategy generation in step S300. The graph neural network receives the dynamically updated graph and outputs the risk score for each node. In a certain production run, the risk score of the "stator insulation resistance" node suddenly rose to 0.92, exceeding the preset threshold of 0.8. The system immediately initiates a reverse traversal algorithm, tracing along the path with the largest edge weight, and finds that the source of the risk points to the "immersion furnace temperature control" node, and further traces it to the "winding tension fluctuation" node. Based on this risk transmission path, the system queries the process specification library, determines that "winding tension" is an adjustable process parameter, and, combined with the historical best practice range, generates a process adjustment strategy: "It is recommended to fine-tune the tension setting of winding machine A from 15N to 14.5N and check the wear of the guide rollers."

[0085] Simultaneously, if a new model of motor process document is introduced on-site, the system calls a multimodal large language model to parse the document. The model reads the "new insulating varnish formulation requirements" and the accompanying "coating thickness diagram" from the text, extracting a triple (head entity: "new insulating varnish", relation: "recommended coating thickness", tail entity: "0.2mm±0.05mm"). After a confidence score higher than 0.9, the triple is injected into the quality knowledge graph, updating the attribute constraints of the relevant nodes. In addition, for this high-risk event, the system inputs the abnormal node subgraph into the large language model. The model combines parameterized knowledge to output a natural language root cause description: "Suspected to be due to excessive tension causing localized thinning of the varnish film, leading to a decrease in insulation performance," and performs consistency verification with the generated risk transmission path to confirm logical consistency.

[0086] Finally, the push and feedback loop of step S400 is executed. The system pushes the risk path visualization interface and adjustment strategy to the operator's industrial tablet terminal. After the operator confirms and executes the adjustment, the insulation resistance pass rate of the next batch returns to the normal level. The system calculates the change in quality indicators before and after execution (e.g., the defect rate decreases from 5% to 0.5%), and sends this change back as a positive feedback signal. The graph neural network uses this feedback signal to update its internal parameters through gradient descent, enhancing the weighted reliability of the "tension-insulation resistance" path, enabling the model to issue an early warning under similar operating conditions in the future.

[0087] It is worth noting that the factory has three production bases, each with sensitive data privacy. Therefore, the updates to the aforementioned model parameters are not performed locally but are incorporated into a federated learning framework. Each base's local client iteratively trains its local model based on its own time-aligned data and spectral features, and uploads the encrypted gradients to the central server for secure aggregation. This achieves collaborative optimization of the global quality prediction model and rapid adaptation to new production lines without sharing the original data.

[0088] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A digitally-based multi-mode production operation assistance and quality control method, characterized in that, Includes the following steps: Acquire multi-source heterogeneous data from a multi-mode production process. The multi-source heterogeneous data includes equipment operating parameters, process data, environmental condition data, and quality inspection data. For each pair of data sources, calculate the mutual information between two data sequences under different delay parameters within a preset delay parameter range. Select the delay parameter that reaches the maximum value of the mutual information as the optimal delay parameter between the pair of data sources. Perform time shift compensation on the two data sequences based on the optimal delay parameter to obtain time-aligned multi-source heterogeneous data. Using production factors as nodes and process flow relationships and quality causal relationships as edges, a quality knowledge graph is constructed based on the time-aligned multi-source heterogeneous data. The production factors include equipment, processes, materials, and quality indicators. The edges include edges representing process flow relationships and edges representing quality causal relationships. The edge weights of the edges representing quality causal relationships are determined based on the mutual information of the connected nodes in the time-aligned data sequence. The quality knowledge graph is dynamically reasoned through a graph neural network to output a quality risk score for each node. Based on the nodes whose quality risk scores exceed a preset threshold, a quality risk transmission path and process adjustment strategy are generated. The quality risk transmission path and the process adjustment strategy are pushed to the operation terminal. The execution result of the process adjustment strategy by the operator is sent back as a feedback signal. The parameters of the graph neural network are updated online based on the feedback signal.

2. The method for digital-based multi-mode production operation assistance and quality control according to claim 1, characterized in that, The feature vectors of the nodes in the quality knowledge graph contain the statistical features of the time-aligned data sequence, including mean, variance, trend slope, and peak interval. The edges of the quality knowledge graph are divided into explicit edges and implicit edges. The explicit edges represent known process flow relationships, and the implicit edges represent the association relationships between node pairs whose mutual information exceeds a preset mutual information threshold.

3. The method for digital-based multi-mode production operation assistance and quality control according to claim 1, characterized in that, The mutual information is calculated as follows: for any two data sequences X and Y, with a delay parameter... Next, calculate the joint probability distribution. With marginal probability distribution and The relative entropy between them, the formula for calculating the relative entropy is: Where X and Y are random variables, and x and y are the specific values ​​that X and Y take, respectively. For delay parameters, The mutual information, expressed in bits, is in the delay parameter. The value below, For delay parameters The joint probability distribution when X takes the value x and Y takes the value y. Let X be the marginal probability distribution when X takes the value x. Let Y be the marginal probability distribution when Y takes the value y. This represents the logarithm to base 2, summing all possible values ​​of x and y, and selecting the value that makes the sum... Reaching the maximum value This serves as the optimal delay parameter.

4. The method for digital-based multi-mode production operation assistance and quality control according to claim 1, characterized in that, The process adjustment strategy is generated based on the adjustable process parameters of each node in the quality risk transmission path; the feedback signal is the change in quality indicators before and after the operator executes the process adjustment strategy; the graph neural network updates its parameters based on the feedback signal after normalization, smoothing filtering and outlier removal preprocessing.

5. The method for digital-based multi-mode production operation assistance and quality control according to claim 2, characterized in that, The quality risk transmission path is generated as follows: starting from the node whose quality risk score exceeds the preset threshold, the path in the quality knowledge graph that satisfies the preset edge weight condition is traversed in reverse until the material node or equipment node is reached, and the sequence of nodes traversed in the reverse traversal is taken as the quality risk transmission path.

6. The method for digital-based multi-mode production operation assistance and quality control according to claim 1, characterized in that, Also includes: The process document data of the multi-mode production process is acquired, and the process document data is semantically parsed using a multimodal large language model to extract process routes, processes, steps, and process parameter constraints. The extraction results are injected into the quality knowledge graph in the form of triples as new nodes and edges. The multimodal large language model receives the text content of the process document and the associated product image data, and evaluates the confidence of the parsed triples. Only triples with a confidence score higher than a preset confidence threshold are injected into the quality knowledge graph.

7. The method for digital-based multi-mode production operation assistance and quality control according to claim 6, characterized in that, When the graph neural network outputs a node whose quality risk score exceeds the preset threshold, the associated data of that node and the neighbor subgraph of that node in the quality knowledge graph are input into the multimodal large language model. The multimodal large language model performs root cause analysis on the quality deviation based on parameterized knowledge, outputs a root cause description and recommended corrective measures in natural language form, and verifies the consistency between the root cause description and the corrective measures and the quality risk transmission path.

8. The method for digital-based multi-mode production operation assistance and quality control according to claim 1, characterized in that, Also includes: Construct a federated learning framework, which includes a central server and local clients deployed in multiple production sites; Each local client iteratively trains a local quality prediction model based on multi-source heterogeneous data aligned to local time. The input of the local quality prediction model is the feature vector of the node in the quality knowledge graph, and the output is the quality risk score. Each local client uploads the gradient parameters obtained from the iterative training to the central server after homomorphic encryption.

9. The method for digital-based multi-mode production operation assistance and quality control according to claim 8, characterized in that, The central server securely aggregates the gradient parameters uploaded by each local client to obtain the parameters of the global quality prediction model, and then distributes the parameters of the global quality prediction model to each local client. Each local client uses the global quality prediction model as the teacher model and a locally deployed lightweight model as the student model. By minimizing the divergence loss between the output distribution of the teacher model and the output distribution of the student model, knowledge transfer from the global quality prediction model to the student model is achieved.

10. The method for digital-based multi-mode production operation assistance and quality control according to claim 9, characterized in that, When a new production line or new product model is introduced, the local client performs rapid fine-tuning of the student model on a small number of local samples based on a model-independent meta-learning framework. The model-independent meta-learning framework enables the student model to quickly adapt to the new data distribution under a small number of samples by calculating task-specific gradients based on support set samples in the inner loop and updating meta-parameters based on query set samples in the outer loop.

Images