Full-process visualized tracing method and system based on production site data

Abstract

This invention relates to the field of production quality control and product traceability technology, and discloses a full-process visual traceability method and system based on production site data. The system includes multi-source heterogeneous data collection and anomaly warning and visual traceability links for the entire production process. The steps are as follows: S1. The strongly coupled operating condition parameter data of each process node in the collected production process are decoupled based on a multivariate adaptive regression model. The cross-coupling influence between multiple parameters is removed, and the independent marginal contribution of each single parameter to the quality result of the corresponding process is quantified. The decoupled independent feature sets of each process traceability node are generated and output. This invention constructs a correlation model between parameters and quality results using the decoupled pure feature data, and uses a logistic regression model to generate an initial dynamic compliance threshold range adapted to the current production scenario. This effectively replaces traditional fixed thresholds, improving the accuracy of threshold generation and scenario adaptability from the modeling source.

Description

Technical Field

[0001] This invention relates to the field of production quality control and product traceability technology, and more specifically, to a full-process visual traceability method and system based on production site data. Background Technology

[0002] Existing production traceability technologies have significant shortcomings in threshold setting. They generally adopt a crude control model that combines manual experience with single statistical values ​​to set fixed thresholds, lacking the ability to adapt to dynamic changes in production scenarios. Such fixed thresholds fail to fully explore the inherent correlation between operating parameters and process quality in historical production data, and do not consider the impact of multi-dimensional scenario factors such as product model switching, material batch changes, equipment operating status fluctuations, and workshop environment changes. This results in a serious disconnect between thresholds and actual production conditions, making them susceptible to interference from factors such as operating condition drift and equipment aging, frequently leading to false alarms and missed alarms. Consequently, they cannot provide a scientific and accurate judgment benchmark for quality control, and are unable to meet the needs of refined and intelligent control in industrial production.

[0003] Existing technologies lack effective dynamic optimization mechanisms for thresholds. Most traceability systems can only achieve one-time static modeling, and once the thresholds are set, they cannot be autonomously adjusted according to real-time changes in the production process. They require frequent manual intervention and calibration, which not only increases the workload of operators but also suffers from problems such as untimely adjustments and low calibration accuracy. At the same time, existing threshold control models have not built a closed-loop feedback mechanism centered on real-time operating conditions, control effectiveness, and finished product quality. They cannot continuously optimize threshold parameters based on key indicators such as false alarm rate, false negative rate, and finished product qualification rate of anomaly warnings. This results in poor long-term system stability, difficulty in improving anomaly identification accuracy, and an inability to achieve efficient and accurate quality control throughout the entire production process.

[0004] Existing production traceability technologies have shortcomings in the connection between data processing and threshold modeling. They fail to effectively decouple strongly coupled operating parameters collected from the production site. The cross-coupling interference between parameters makes it impossible to accurately quantify the independent impact of individual parameters on process quality. Consequently, the parameter-quality correlation model built based on such coupled data becomes distorted, and the accuracy of threshold generation is greatly reduced. This further exacerbates the problems of poor threshold adaptability and frequent false alarms and missed alarms. As a result, it cannot provide reliable data support for subsequent anomaly tracing and root cause localization, thus restricting the application effect and promotion value of production traceability technology. Summary of the Invention

[0005] This invention provides a full-process visual traceability method and system based on production site data, which solves the technical problems mentioned in the background art.

[0006] This invention provides a full-process visual traceability method based on production site data, including multi-source heterogeneous data collection and anomaly early warning and visual traceability links throughout the entire production process, with the following steps: S1. Decouple the strongly coupled operating condition parameter data of each process node in the entire production process based on the multivariate adaptive regression model, remove the cross-coupling influence between multiple parameters, quantify the independent marginal contribution of each single parameter to the quality result of the corresponding process, and generate and output the decoupled independent feature set of each process traceability node. S2. Based on the decoupled independent feature set obtained in step S1, an adaptive modeling framework integrating reinforcement learning and logistic regression is constructed. Using historical working condition data, material batch data, and quality inspection data corresponding to each process traceability node as training aids, the mapping relationship between the decoupled independent features and the process quality results is first fitted through a logistic regression model to generate the initial dynamic compliance threshold range for each traceability node. Then, through a preset reinforcement learning agent, the initial dynamic compliance threshold range is iteratively optimized online using real-time working conditions, threshold control effects, and finished product quality results as closed-loop feedback, and a real-time dynamic compliance threshold range adapted to the current production scenario is output. S3. Based on the real-time dynamic compliance threshold range of step S2, online compliance judgment is performed on the real-time operation data of each process traceability node, abnormal traceability nodes are identified and the corresponding compliance risk level is marked, and abnormal traceability signals are generated synchronously. S4. Based on the abnormal traceability signal in step S3, complete the root cause location of the entire abnormal traceability chain, and simultaneously realize the linkage visualization display of the entire production process traceability data, real-time dynamic compliance threshold range, compliance risk level and abnormal traceability chain on the visualization interface.

[0007] Furthermore, step S1 includes the following steps: S1.1. For the collected strongly coupled working condition parameter data of each process node, align the time dimension with the production cycle time of the corresponding process, remove invalid null values ​​and abnormal jump data, and generate a standardized coupled parameter dataset. S1.2 The quality inspection result of the current process is the output dependent variable, and the parameters of each working condition in the standardized coupled parameter dataset are the input independent variables to construct a multivariate adaptive regression equation with an adaptive regularization term. S1.3 Solve the partial derivatives of the multivariate adaptive regression equation, remove the cross-coupling terms between the input independent variables, calculate the pure influence coefficient of each input independent variable on the output dependent variable, and quantify the independent marginal contribution of the corresponding single parameter to the process quality result based on the pure influence coefficient. S1.4 Sort each single parameter based on independent marginal contribution, remove redundant parameters with contribution below the preset threshold, retain effective core parameters, and generate decoupled independent feature sets for the corresponding process traceability nodes.

[0008] Furthermore, step S1 also includes the following sub-steps: S1.5 Decouple each single-parameter feature in the independent feature set and bind a unique, tamper-proof, traceable time anchor point, process identity identifier, and production batch identifier that are strongly associated with the corresponding process production cycle time. S1.6 Perform feature consistency and integrity checks on the decoupled independent feature sets of the bound traceability anchors. After passing the checks, store them in the distributed traceability database for subsequent threshold modeling and full-process traceability calls.

[0009] Furthermore, in step S2, the initial dynamic compliance threshold range for each traceability node is generated using a logistic regression model, including the following steps: S2.1. The historical decoupled independent feature set of the corresponding process traceability node is used as the input sample. The process quality qualified / unqualified result of the corresponding historical batch is used as the binary label. The corresponding historical working condition data, material batch data, and equipment operation data are matched to construct a standardized training sample set for the logistic regression model. S2.2. Supervised training of the logistic regression model is performed based on the standardized training sample set. The nonlinear mapping relationship between each single parameter feature in the decoupled independent feature set and the process quality result is fitted to complete the model convergence verification and generalization accuracy verification. S2.3 Based on the preset confidence level, and combined with the product model, material batch, equipment status, and environmental parameters of the current production scenario, the trained logistic regression model outputs the upper and lower limits of the initial dynamic compliance threshold range for the corresponding traceability node.

[0010] Furthermore, in step S2, the construction of the preset reinforcement learning agent specifically includes the following sub-steps: S2.4. Using the real-time working conditions of the entire production process as the intelligent agent's interaction environment, a multi-dimensional state space is constructed. The state vector of the state space includes the decoupled independent feature set of the current process, the initial dynamic compliance threshold range, real-time working condition parameters, real-time quality inspection results, equipment operating status, and compliance risk level. S2.5 Construct the executable action space of the intelligent agent. The action space includes fine-tuning actions of the upper and lower limits of the initial dynamic compliance threshold range, adjustment actions of the weights of each feature in the decoupled independent feature set, and regularization coefficient optimization actions of the logistic regression model. Simultaneously set the physical boundary constraints and adjustment step size rules for each action. S2.6 The core optimization targets for threshold control are false alarm rate, false negative rate, and finished product qualification rate. A reward function is set: when the agent performs an action, if the false alarm rate and false negative rate of threshold control decrease and the finished product qualification rate increases, a positive reward is given; if the false alarm rate and false negative rate increase or the finished product qualification rate decreases, a negative penalty is given.

[0011] Furthermore, in step S2, the initial dynamic compliance threshold range is optimized online through a reinforcement learning agent, specifically including the following sub-steps: S2.7. Collect the real-time state vector of the current production scenario according to the preset iteration cycle, and input it into the reinforcement learning agent after standardization processing. S2.8 The agent, based on the current state vector, aims to maximize the cumulative reward and outputs the optimal action through the policy network to complete the online update of the initial dynamic compliance threshold range. S2.9. Collect the real-time control effect and finished product quality results after the threshold update, calculate the corresponding reward value and feed it back to the agent to complete one iteration cycle; continue to execute the iteration cycle until the control effect of the threshold range reaches the preset stable standard, and output the final real-time dynamic compliance threshold range adapted to the current scenario.

[0012] Furthermore, step S3 specifically includes the following sub-steps: S3.1 Match and compare the real-time operation data of each process traceability node with the real-time dynamic compliance threshold range of the corresponding node to determine whether the real-time operation data is within the threshold range. S3.2 Based on the magnitude of data deviation, duration of deviation, and independent marginal contribution of corresponding parameters, traceability nodes are divided into four compliance risk levels: no risk, low risk, medium risk, and high risk, and each node is marked with a corresponding level. S3.3 For traceability nodes marked as medium risk or above, they are determined to be abnormal traceability nodes. An abnormal traceability signal containing the abnormal node's identity, out-of-tolerance parameter information, risk level, and traceability time anchor point is generated and pushed to the visual traceability module.

[0013] Furthermore, in step S4, the root cause localization of the entire process anomaly tracing chain is completed, specifically including the following sub-steps: S4.1 Analyze the abnormal traceability signal, lock the production batch, process link and time window corresponding to the abnormal traceability node, and determine the boundary range of the whole process traceability; S4.2 In the distributed traceability database, retrieve the decoupled independent feature set, historical threshold range, real-time operation data, quality inspection data and operation record data of all process nodes within the locked traceability range; S4.3. Based on the independent marginal contribution of each parameter, quantify the impact weight of each process node on the abnormal result within the traceability range, locate the core root cause node and the corresponding parameter deviation, and reconstruct the entire abnormal impact chain from the root cause node to the abnormal node.

[0014] Furthermore, step S4 includes the following steps of linked visualization display: S4.4 Render a directed process topology diagram of the entire production process in the visualization interface, with a single process traceability node as the smallest visualization unit, and display the basic attribute information of each node accordingly. S4.5. Use floating color bands to render the real-time dynamic compliance threshold range of each process traceability node, and use gradient colors that correspond one-to-one with the compliance risk level to visually label each process node. S4.6 After receiving the abnormal traceability signal, the abnormal impact link and core root cause node are automatically highlighted in the process topology diagram, supporting users to drill down and trace back up from batch level to single parameter millisecond level traceability data. S4.7 Based on user interaction, the full-process traceability data, threshold change curves, risk level change records, quality inspection results and operation records within the corresponding traceability range are displayed synchronously on a unified timeline of the same visualization interface. It also includes the optimization steps for full-process traceability and closed-loop management: S4.8 Based on the core root cause nodes and parameter deviations of the location, a forward simulation model of the entire process parameters is constructed. The full process quality results after parameter adjustment are simulated to verify the accuracy of the root cause location. Based on the simulation verification results, S4.9 outputs parameter optimization suggestions for the corresponding process nodes. Simultaneously, it supplements the root cause data, optimization results, and quality feedback data of this anomaly tracing into the training sample set of the logistic regression model, completes the offline iterative update of the model, and forms a closed loop of full-process control for tracing, positioning, simulation, verification, optimization, and model update.

[0015] Furthermore, the end-to-end visual traceability system based on production site data includes multi-source heterogeneous data acquisition and anomaly early warning and visual traceability links throughout the entire production process. Its features are as follows: For the strongly coupled operating condition parameter data of each process node in the entire production process, a multivariate adaptive regression model is used to decouple the parameters, remove the cross-coupling influence between multiple parameters, quantify the independent marginal contribution of each single parameter to the quality result of the corresponding process, and generate and output the decoupled independent feature set of each process traceability node. An adaptive modeling framework integrating reinforcement learning and logistic regression is constructed. Historical operating data, material batch data, and quality inspection data corresponding to each traceability node are used as training aids. First, the mapping relationship between decoupled independent features and process quality results is fitted by a logistic regression model to generate the initial dynamic compliance threshold range for each traceability node. Then, through a pre-set reinforcement learning agent, the initial dynamic compliance threshold range is iteratively optimized online with real-time operating conditions, threshold control effects, and finished product quality results as closed-loop feedback, and the real-time dynamic compliance threshold range adapted to the current production scenario is output. Online compliance assessment is performed on the real-time operational data of each process traceability node, abnormal traceability nodes are identified and their corresponding compliance risk levels are marked, and abnormal traceability signals are generated simultaneously. Complete the root cause location of the entire process of anomaly tracing, and simultaneously realize the linked visualization display of the entire production process traceability data, real-time dynamic compliance threshold range, compliance risk level and anomaly tracing link on the visual interface.

[0016] The beneficial effects of this invention are as follows: This invention constructs a correlation model between parameters and quality results using decoupled, pure feature data, and generates an initial dynamic compliance threshold range adapted to the current production scenario using a logistic regression model. This effectively replaces traditional fixed thresholds, improving the accuracy of threshold generation and scenario adaptability from the modeling source. This approach abandons the traditional extensive mode of setting thresholds based on manual experience and single statistical values. Instead, it uses the quality of the process as the learning target, fully learning the inherent laws between parameters and quality in historical data. Combining multi-dimensional scenario information such as current product model, material batch, equipment status, and environmental conditions, it outputs a dynamic threshold range with statistical confidence. This makes the threshold no longer a static fixed value, but an adaptive benchmark strongly correlated with the production scenario. This solves the long-standing problems of poor adaptability, disconnect from actual working conditions, susceptibility to working condition drift, and frequent false alarms and missed alarms caused by traditional fixed thresholds. This makes the quality control benchmark more scientific, reasonable, and highly consistent with the actual production status on site.

[0017] This invention introduces a reinforcement learning agent to perform online self-iterative optimization of the initial dynamic compliance threshold range. Using real-time operating conditions, threshold control effectiveness, and finished product quality as closed-loop feedback, it constructs an intelligent threshold control mechanism that can learn autonomously and continuously evolve. The agent optimizes the false alarm rate, false negative rate, and finished product pass rate. Under the premise of meeting industrial physical boundary constraints, it automatically completes fine-tuning of the upper and lower limits of the threshold, updating feature weights, and optimizing the model regularization coefficients. This enables the threshold range to continuously and adaptively correct itself as the production environment fluctuates, material batches change, equipment performance degrades, and process parameter drifts, without requiring frequent manual intervention or calibration. This mechanism upgrades the system from a one-time static modeling to a long-term online iterative evolution dynamic control mode, further reducing the false alarm and false negative rates, improving anomaly identification accuracy and long-term system stability, and significantly enhancing the level of intelligent quality control throughout the entire production process. Attached Figure Description

[0018] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation

[0019] 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, features described in some examples may be combined in other examples.

[0020] like Figure 1 As shown, a full-process visual traceability method based on production site data includes multi-source heterogeneous data collection and anomaly early warning and visual traceability stages throughout the entire production process. The steps are as follows: S1. Decouple the strongly coupled operating condition parameter data of each process node in the entire production process based on the multivariate adaptive regression model, remove the cross-coupling influence between multiple parameters, quantify the independent marginal contribution of each single parameter to the quality result of the corresponding process, and generate and output the decoupled independent feature set of each process traceability node. S2. Based on the decoupled independent feature set obtained in step S1, an adaptive modeling framework integrating reinforcement learning and logistic regression is constructed. Using historical working condition data, material batch data, and quality inspection data corresponding to each process traceability node as training aids, the mapping relationship between the decoupled independent features and the process quality results is first fitted through a logistic regression model to generate the initial dynamic compliance threshold range for each traceability node. Then, through a preset reinforcement learning agent, the initial dynamic compliance threshold range is iteratively optimized online using real-time working conditions, threshold control effects, and finished product quality results as closed-loop feedback, and a real-time dynamic compliance threshold range adapted to the current production scenario is output. S3. Based on the real-time dynamic compliance threshold range of step S2, online compliance judgment is performed on the real-time operation data of each process traceability node, abnormal traceability nodes are identified and the corresponding compliance risk level is marked, and abnormal traceability signals are generated synchronously. S4. Based on the abnormal traceability signal in step S3, complete the root cause location of the entire abnormal traceability chain, and simultaneously realize the linkage visualization display of the entire production process traceability data, real-time dynamic compliance threshold range, compliance risk level and abnormal traceability chain on the visualization interface.

[0021] Specifically, this method can be applied to various industrial production scenarios such as discrete manufacturing and process manufacturing, covering the entire production lifecycle from raw material entry, process processing, finished product testing, and finished product delivery. The core is to solve the industry pain points of traditional production traceability technology, such as parameter coupling interference, poor threshold adaptability, insufficient traceability accuracy, and lack of control loop, through a complete process of data decoupling, dynamic threshold modeling, hierarchical early warning, full-link traceability and visualization linkage.

[0022] Before implementation, this method first collects multi-source heterogeneous data from the entire production process. The collection process covers all data sources related to product quality, equipment operation, and process flow on the production site, including but not limited to real-time operating data of production line PLC equipment, quality inspection data of quality inspection equipment, production plan and process flow data of MES system, material batch data of ERP system, workshop environmental data of environmental sensors, and personnel operation record data of on-site operation terminals. For the differences in collection frequency and data format of different data sources, a unified standardized interface is used to aggregate the data, providing a complete original data foundation for subsequent decoupling processing, modeling analysis, and traceability.

[0023] The four core steps of this method are executed sequentially, with the output of each preceding step serving as the input for subsequent steps, forming a tightly linked technical loop. Step S1 is the core foundation of the entire method, responsible for refining the raw collected data, eliminating coupling interference between parameters, and providing interference-free feature data for subsequent modeling. Step S2 is responsible for building a dynamic compliance threshold system adapted to production scenarios, overcoming the technical limitations of traditional fixed thresholds. Step S3 is the bridge connecting threshold control and traceability, responsible for real-time compliance judgment and anomaly triggering. Step S4 is the implementation output of this method, responsible for anomaly root cause localization, visualization, and full-process closed-loop optimization.

[0024] Multi-source heterogeneous data acquisition throughout the entire production process refers to the process of collecting and aggregating production-related data from different data sources, in different formats, and at different collection frequencies, covering the entire production lifecycle from raw material entry to finished product delivery. The data sources collected include, but are not limited to, production line PLC equipment, quality inspection instruments, MES production management system, ERP enterprise resource management system, environmental sensors, and personnel operation terminals. The collected data content covers six major categories: operating conditions, quality inspection, material flow, process planning, personnel operation, and environmental status, which form the data foundation for all aspects of this method.

[0025] Strongly coupled operating condition parameter data refers to operating condition data collected in the production process where multiple parameters have strong cross-influence relationships. Changes in these parameters will affect each other, and it is impossible to directly distinguish the independent impact of a single parameter on product quality. For example, in the injection molding process, the barrel temperature and injection pressure. Temperature changes will affect the fluidity of the melt, which in turn will cause changes in the impact of pressure on molding quality. The two have a strong coupling relationship and are the core object of the decoupling process in step S1 of this method.

[0026] The multivariate adaptive regression model refers to the adaptively adjustable regression model used for parameter decoupling in this method. The multivariate adaptive regression model takes the process quality result as the output and multiple coupled working condition parameters as the input. The model characteristics can be adjusted through adaptive regularization terms, and the coupling effect between parameters can be separated by solving the partial derivatives. The independent influence of each parameter on the quality result can be accurately quantified. It is the core basic model for feature purification in this method.

[0027] The decoupled independent feature set refers to the feature set that has been decoupled by the multivariate adaptive regression model, eliminating the cross-coupling effects between parameters and retaining the core parameters that have a significant impact on process quality. Each feature in the set corresponds to a working condition parameter, which can truly reflect the independent impact of the parameter on process quality. Moreover, each feature is bound to a unique traceability identifier, which is the core data foundation for subsequent threshold modeling and root cause tracing.

[0028] The logistic regression model refers to the binary classification model used in this method to generate the initial dynamic compliance threshold range. The logistic regression model takes a decoupled independent feature set as input and the process quality as qualified or unqualified as the classification label. It can fit the nonlinear mapping relationship between the parameter features and the quality qualification probability, and output the compliance threshold range of the corresponding parameters. It has strong interpretability and fully meets the traceability and auditability requirements of industrial production.

[0029] The initial dynamic compliance threshold range refers to the upper and lower limits of compliance for each core parameter output by the trained logistic regression model, combined with information such as product model, material batch, and equipment status in the current production scenario. The initial dynamic compliance threshold range differs from the traditional fixed threshold and can be automatically adjusted according to changes in the production scenario, which is the basis for subsequent online optimization through reinforcement learning.

[0030] The real-time dynamic compliance threshold range refers to the upper and lower limits of parameter compliance that are finally output after online self-iterative optimization by the reinforcement learning agent and are fully adapted to the current real-time production scenario. The real-time dynamic compliance threshold range will be continuously iterated and optimized as the production process progresses, and can adapt to various scenario changes such as operating condition drift, equipment aging, and material changes. It is the core benchmark for online compliance judgment in this method.

[0031] The compliance risk level refers to the risk classification of process traceability nodes based on the degree of parameter deviation and its impact on quality. In this method, it is divided into four levels: no risk, low risk, medium risk, and high risk. Different levels correspond to different control strategies, which can realize graded early warning and precise control, and avoid the interference of ineffective early warnings on production.

[0032] Anomaly tracing signal refers to an instruction signal automatically generated for anomaly tracing nodes of medium risk or above, used to trigger full-process tracing. The signal contains core information such as the identity of the anomaly node, out-of-tolerance parameter information, risk level, and tracing time anchor point, and is the trigger condition for starting full-process root cause tracing.

[0033] Production process traceability data refers to all data that can be used for product lifecycle traceability, including but not limited to decoupled independent feature sets of each process node, historical and real-time threshold ranges, operating condition data, quality inspection data, material batch data, personnel operation records, equipment operation data, threshold control records, and abnormal handling records. It covers the entire process from raw materials to finished products and is the core content for root cause localization and visualization.

[0034] In one embodiment of the present invention, step S1 includes the following steps: S1.1. For the collected strongly coupled working condition parameter data of each process node, align the time dimension with the production cycle time of the corresponding process, remove invalid null values ​​and abnormal jump data, and generate a standardized coupled parameter dataset. S1.2 The quality inspection result of the current process is the output dependent variable, and the parameters of each working condition in the standardized coupled parameter dataset are the input independent variables to construct a multivariate adaptive regression equation with an adaptive regularization term. S1.3 Solve the partial derivatives of the multivariate adaptive regression equation, remove the cross-coupling terms between the input independent variables, calculate the pure influence coefficient of each input independent variable on the output dependent variable, and quantify the independent marginal contribution of the corresponding single parameter to the process quality result based on the pure influence coefficient. S1.4 Sort each single parameter based on independent marginal contribution, remove redundant parameters with contribution below the preset threshold, retain effective core parameters, and generate decoupled independent feature sets for the corresponding process traceability nodes.

[0035] Specifically, steps S1.1 to S1.4 in this embodiment are a complete decoupling process for strongly coupled operating condition parameter data. The four sub-steps are executed sequentially according to the order of data processing, and finally output a decoupled independent feature set that can be used for subsequent modeling.

[0036] During the execution of step S1.1, the production cycle benchmark for the corresponding process is first defined. The production cycle refers to the standard cycle required for the process to complete the processing of a single qualified workpiece. It is the core time benchmark for process flow in the entire production process and the core anchor point for solving the problem of time asynchrony of multi-source data. For the strongly coupled working condition parameter data collected from different data sources, the timestamps of all data are uniformly aligned using the production cycle of the process as the smallest time unit. Data with different collection frequencies are uniformly mapped to the same time dimension to ensure that all working condition parameter data within the same production cycle corresponds to the processing of the same workpiece, avoiding the problem of data time sequence disorder. After completing the time alignment, the dataset is cleaned to remove invalid null values, including missing data caused by sensor disconnection or network interruption. At the same time, abnormal jump data is identified and removed, including non-processing state data at the moment of equipment start-up and shutdown, and abnormal data exceeding physical limits caused by sensor failure. Finally, a standardized coupled parameter dataset with uniform format, accurate time sequence, and no invalid data is generated.

[0037] Step S1.2 involves constructing a decoupling regression model based on the standardized dataset, establishing a model framework that establishes the correlation between process quality results and operating parameters. During execution, the output and input dimensions of the model are first determined. The output dependent variable uses the quality inspection results of the corresponding workpiece for that process. These quality inspection results are quantifiable quality indicators, such as dimensional tolerances, geometric tolerances, appearance defect levels, and performance test results, which can be directly used to determine whether the process is qualified. The input independent variables use all operating parameters from the standardized coupling parameter dataset. These operating parameters are real-time parameters collected during the process that may affect product quality, such as spindle speed, feed rate, and depth of cut in machining; barrel temperature, injection pressure, and holding time in injection molding; and welding current, welding voltage, and welding speed in welding. These parameters generally have strong cross-coupling effects and are the core objects of subsequent decoupling. When constructing the model, an adaptive regularization term is added to balance the model's fitting accuracy and generalization ability, avoiding overfitting and allowing the model to automatically adjust the regularization strength based on changes in subsequent input data, adapting to changes in different production scenarios.

[0038] Step S1.3 is the core of the entire decoupling process, used to isolate the cross-coupling effects between parameters and quantify the independent influence of each parameter. During execution, partial derivatives are calculated on the regression equation constructed in step S1.2. Through the partial derivative calculation process, the influence of each input independent variable on the output dependent variable is isolated from the mixed influence of multiple parameters. The pure influence coefficient of each working condition parameter on the process quality result is calculated. The pure influence coefficient can directly reflect the degree of independent influence of the parameter on the process quality when it is not disturbed by other parameters. Based on the calculated pure influence coefficient, the independent marginal contribution of each single parameter to the process quality result is further quantified. The independent marginal contribution is presented in the form of a percentage, which can intuitively reflect the influence weight of each parameter on the process quality. The higher the contribution, the greater the impact of the change of the parameter on the process quality result.

[0039] Step S1.4 is used to complete the feature selection and feature set generation after decoupling. During the execution process, based on the independent marginal contribution calculated in step S1.3, all working condition parameters are sorted from high to low according to their contribution, clarifying the priority of each parameter's impact on process quality. Then, a preset threshold is set, which can be adjusted according to the accuracy requirements of the production scenario. Redundant parameters with independent marginal contribution below the preset threshold are eliminated. Redundant parameters have minimal impact on the process quality results, and retaining them will increase the computational load of subsequent models without improving model accuracy. Finally, the effective core parameters with independent marginal contribution that meet the requirements are retained, and the decoupled data corresponding to these parameters are combined to generate the decoupled independent feature set of the corresponding process traceability node. The decoupled independent feature set completely eliminates the cross-coupling interference between parameters and can truly reflect the independent impact of each parameter on process quality.

[0040] In one embodiment of the present invention, step S1 further includes the following sub-steps: S1.5 Decouple each single-parameter feature in the independent feature set and bind a unique, tamper-proof, traceable time anchor point, process identity identifier, and production batch identifier that are strongly associated with the corresponding process production cycle time. S1.6 Perform feature consistency and integrity checks on the decoupled independent feature sets of the bound traceability anchors. After passing the checks, store them in the distributed traceability database for subsequent threshold modeling and full-process traceability calls.

[0041] Specifically, steps S1.5 to S1.6 of this embodiment are traceability attribute binding and storage management processes for decoupled independent feature sets, which are used to ensure that the decoupled feature data has traceability, immutability and callability, and provide a data foundation for subsequent full-process traceability.

[0042] Step S1.5 involves binding a unique traceability identifier to each decoupled feature data, achieving the effect of feature-based traceability. During execution, for each single-parameter feature in the decoupled independent feature set, three sets of unique traceability identifiers are bound. The first set is a traceability time anchor point strongly correlated with the production cycle time of the corresponding process. The traceability time anchor point corresponds exactly to the production cycle time used for data alignment in step S1.1, which can accurately locate the workpiece, processing time, and process node corresponding to the feature data, ensuring the time sequence of the data is traceable. The second set is a process identifier, which includes the production line number, process number, and equipment number of the process, which can accurately locate the production location and equipment corresponding to the feature data. The third set is a production batch identifier, which includes the material batch, production work order number, and product model corresponding to the workpiece, which can accurately locate the production batch and product information corresponding to the feature data. The combination of the three sets of identifiers forms a unique and tamper-proof traceability identifier, which is strongly bound to the corresponding feature data. This ensures that in the subsequent traceability process, the corresponding feature data can be quickly located through any set of identifiers, while avoiding the problem of untraceability after data tampering.

[0043] Step S1.6 involves verifying and managing the storage of feature data bound to traceability identifiers. During execution, the decoupled independent feature sets bound to traceability anchors undergo dual verification. The first verification is feature consistency, checking whether the parameter dimensions and data range of the feature set are consistent with the preset parameter boundaries of the process to avoid data corruption and binding errors. The second verification is integrity, checking whether all core parameters in the feature set have been bound to traceability identifiers and whether there are any missing data or identifiers. Only feature sets that pass both verifications are stored in the distributed traceability database. The distributed traceability database uses a multi-node distributed deployment, enabling secure storage and rapid retrieval of massive amounts of production data. The database also features tiered access control, allowing operators with different roles to access data only within their corresponding access permissions, ensuring data security and compliance. Feature sets stored in the database are simultaneously indexed based on traceability identifiers. Subsequent threshold modeling and root cause tracing processes can quickly retrieve the corresponding feature data through the indexed directory without traversing the entire database, significantly improving data retrieval efficiency.

[0044] In one embodiment of the present invention, step S2, generating the initial dynamic compliance threshold range for each traceability node using a logistic regression model, includes the following steps: S2.1. The historical decoupled independent feature set of the corresponding process traceability node is used as the input sample. The process quality qualified / unqualified result of the corresponding historical batch is used as the binary label. The corresponding historical working condition data, material batch data, and equipment operation data are matched to construct a standardized training sample set for the logistic regression model. S2.2. Supervised training of the logistic regression model is performed based on the standardized training sample set. The nonlinear mapping relationship between each single parameter feature in the decoupled independent feature set and the process quality result is fitted to complete the model convergence verification and generalization accuracy verification. S2.3 Based on the preset confidence level, and combined with the product model, material batch, equipment status, and environmental parameters of the current production scenario, the trained logistic regression model outputs the upper and lower limits of the initial dynamic compliance threshold range for the corresponding traceability node.

[0045] Specifically, steps S2.1 to S2.3 in this embodiment are a complete process of generating an initial dynamic compliance threshold range through a logistic regression model. Based on the decoupled clean feature data, a correlation model between parameters and quality results is constructed to generate an initial threshold that is adapted to the current production scenario, replacing the traditional fixed threshold.

[0046] Step S2.1 involves constructing a standardized training sample set for the logistic regression model, providing high-quality foundational data for model training. During execution, the historical decoupled independent feature set of the corresponding process traceability node is retrieved from the distributed traceability database. This set serves as the input sample for model training. The input samples have undergone coupling interference removal, accurately reflecting the independent impact of each parameter on the quality results, thus avoiding model distortion caused by coupled data. Subsequently, a corresponding binary label is matched for each input sample. The label indicates whether the process quality of the corresponding historical batch is qualified or unqualified, with qualified results marked as positive and unqualified results as negative. After matching samples with labels, corresponding historical operating data, material batch data, and equipment operation data are added to each sample, including equipment runtime, material supplier information, workshop environmental parameters, and production personnel information during the batch production. This data serves as auxiliary features for subsequent scenario-based adaptation of threshold ranges. After all sample data is matched, it undergoes standardization processing to unify the data format and magnitude, ultimately constructing a standardized training sample set for the logistic regression model.

[0047] The core of step S2.2 is to complete the training and validation of the logistic regression model and establish the mapping relationship between decoupled features and process quality results. During execution, the standardized training sample set is first divided into a training set and a validation set. The training set is used for model fitting training, and the validation set is used for model accuracy validation. Based on the training set, the logistic regression model undergoes supervised training. The model automatically fits the nonlinear mapping relationship between each single-parameter feature in the decoupled independent feature set and the process quality pass / fail result, clarifying the impact of each parameter's numerical change on the probability of quality pass. After the initial training of the model is completed, the model undergoes convergence verification and generalization accuracy verification using the validation set. Convergence verification confirms that the model's training process is stable and there will be no accuracy fluctuations. Generalization accuracy verification confirms that the model also has stable prediction accuracy for new samples not involved in training, avoiding overfitting. Only models that pass both verifications will be used for subsequent threshold interval generation.

[0048] Step S2.3 generates an initial dynamic compliance threshold range adapted to the current scenario based on the trained logistic regression model. During execution, a preset confidence level is first set according to the quality control requirements of industrial production. This confidence level can be adjusted according to the product's quality grade requirements; for example, a 99% confidence level can be set for high-precision products, and a 95% confidence level for conventional products. The confidence level represents the probability that the process quality result is qualified when the parameter is within this threshold range. Then, combining specific information from the current production scenario, including the product model being produced, the batch of materials used, the operating status of the equipment, and the environmental parameters of the workshop, this information is input into the trained logistic regression model. The model will then output the initial dynamic compliance threshold range for each core parameter of the corresponding traceability node, specifying the upper and lower limits for each parameter. This initial dynamic compliance threshold range differs from traditional fixed thresholds; it automatically adjusts according to changes in the current production scenario. For example, when material batches are changed or equipment ages and wears out, the threshold range automatically adapts without manual adjustment, solving the problem of poor adaptability of traditional fixed thresholds.

[0049] In one embodiment of the present invention, step S2, the construction of the preset reinforcement learning agent, specifically includes the following sub-steps: S2.4. Using the real-time working conditions of the entire production process as the intelligent agent's interaction environment, a multi-dimensional state space is constructed. The state vector of the state space includes the decoupled independent feature set of the current process, the initial dynamic compliance threshold range, real-time working condition parameters, real-time quality inspection results, equipment operating status, and compliance risk level. S2.5 Construct the executable action space of the intelligent agent. The action space includes fine-tuning actions of the upper and lower limits of the initial dynamic compliance threshold range, adjustment actions of the weights of each feature in the decoupled independent feature set, and regularization coefficient optimization actions of the logistic regression model. Simultaneously set the physical boundary constraints and adjustment step size rules for each action. S2.6 The core optimization targets for threshold control are false alarm rate, false negative rate, and finished product qualification rate. A reward function is set: when the agent performs an action, if the false alarm rate and false negative rate of threshold control decrease and the finished product qualification rate increases, a positive reward is given; if the false alarm rate and false negative rate increase or the finished product qualification rate decreases, a negative penalty is given.

[0050] Specifically, steps S2.4 to S2.6 of this embodiment are a complete process of customizing and constructing a reinforcement learning agent for this production traceability scenario. This process is to build an agent framework with autonomous decision-making capabilities for online optimization of the initial dynamic compliance threshold range, and to achieve real-time adaptive optimization of the threshold range.

[0051] Step S2.4 involves constructing the agent's interactive environment and multi-dimensional state space to provide complete environmental information for the agent's decision-making. During execution, the real-time operating conditions of the entire production process are used as the agent's interactive environment. This environment includes all dynamic factors that may affect the threshold control effect during the production process. The agent can obtain various information in this environment in real time, and can adjust the threshold range in the environment by executing actions and receive feedback on the control effect from the environment. Based on this interactive environment, a multi-dimensional state space is constructed. Each state vector in the state space fully contains the core information of the current production scenario, specifically including the decoupled independent feature set of the current process, the currently effective initial dynamic compliance threshold range, the real-time collected operating parameters, the real-time quality inspection results of the corresponding process, the operating status of the production equipment, and the compliance risk level of each process node. This state vector is updated in real time as the production process progresses, ensuring that the agent always makes decisions based on the latest production scenario information, avoiding the problem of decisions being out of touch with the actual scenario.

[0052] Step S2.5 involves constructing the executable action space of the intelligent agent and setting constraint rules for the actions to ensure that the agent's decisions comply with the physical boundaries and safety requirements of industrial production. During execution, based on the actual needs of threshold optimization, three types of executable action spaces are constructed. The first type is the fine-tuning action of the upper and lower limits of the initial dynamic compliance threshold range. The agent can make small adjustments to the upper and lower limits of the threshold range according to changes in the scenario to adapt to real-time changes in the working conditions. The second type is the adjustment action of the weights of each feature in the decoupled independent feature set. The agent can adjust the weights of corresponding features in the model according to the real-time changes in the impact of different parameters on the quality results, thereby improving the model's performance. Adaptability; the third category is the optimization of the regularization coefficient of the logistic regression model. The agent can adjust the regularization coefficient according to the real-time prediction accuracy of the model to balance the fitting accuracy and generalization ability of the model. While constructing the action space, corresponding physical boundary constraints and adjustment step size rules are set for each action. The physical boundary constraints are used to limit the execution range of the action. For example, the adjustment of the threshold range cannot exceed the physical limits of the equipment parameters to avoid thresholds that do not meet production safety. The adjustment step size rules are used to limit the adjustment range of each action to avoid production instability caused by large fluctuations in the threshold, and to ensure that all decisions of the agent meet the safety requirements of industrial production.

[0053] Step S2.6 involves setting the reward function for the agent, clarifying its optimization goals, and guiding it to iterate towards improving the effectiveness of threshold control. During execution, the three core control indicators for industrial production are used as optimization targets: false alarm rate, false negative rate, and finished product pass rate. The false alarm rate refers to the proportion of parameters that would not actually cause quality defects but are judged as abnormal by the threshold; the false negative rate refers to the proportion of parameters that would actually cause quality defects but are judged as normal by the threshold; and the finished product pass rate refers to the proportion of final product quality that meets the standards. Based on these three core indicators... The system sets a reward function for each indicator. When an agent performs an action, if the false positive rate and false negative rate decrease while the product qualification rate increases, the agent receives a positive reward. The greater the improvement in the indicator, the higher the positive reward. Conversely, if the false positive rate and false negative rate increase, or the product qualification rate decreases, the agent receives a negative penalty. The greater the deterioration in the indicator, the higher the negative penalty. By setting the reward function, the agent can be guided to continuously optimize the threshold range, ultimately achieving the control objective of minimizing the false positive rate and false negative rate while maximizing the product qualification rate.

[0054] In one embodiment of the present invention, step S2, which involves online self-iterative optimization of the initial dynamic compliance threshold range by a reinforcement learning agent, specifically includes the following sub-steps: S2.7. Collect the real-time state vector of the current production scenario according to the preset iteration cycle, and input it into the reinforcement learning agent after standardization processing. S2.8 The agent, based on the current state vector, aims to maximize the cumulative reward and outputs the optimal action through the policy network to complete the online update of the initial dynamic compliance threshold range. S2.9. Collect the real-time control effect and finished product quality results after the threshold update, calculate the corresponding reward value and feed it back to the agent to complete one iteration cycle; continue to execute the iteration cycle until the control effect of the threshold range reaches the preset stable standard, and output the final real-time dynamic compliance threshold range adapted to the current scenario.

[0055] Specifically, steps S2.7 to S2.9 in this embodiment are the complete execution process of the reinforcement learning agent performing online self-iterative optimization of the initial dynamic compliance threshold range. Through real-time closed-loop feedback, the threshold range continuously adapts to the dynamic changes in the production scenario, and outputs the final real-time dynamic compliance threshold range.

[0056] Step S2.7 involves collecting and inputting real-time state information to provide the latest environmental information for the iterative optimization of the agent. During execution, a preset iteration cycle is first set based on the rate of change and control precision requirements of the production scenario. The iteration cycle can be adjusted according to the production scenario; for example, a shorter iteration cycle can be set for precision machining processes with rapidly changing conditions, while a longer iteration cycle can be set for routine processes with stable conditions, ensuring a balance between the real-time performance of optimization and the stability of production. At the beginning of each iteration cycle, real-time state vectors of the current production scenario are collected from various data sources on the production site. The dimension of the state vectors is completely consistent with the state space set in step S2.4. After collection, the real-time state vectors are standardized to unify the data format and magnitude, ensuring complete matching with the input requirements of the agent. Subsequently, the standardized real-time state vectors are input into the reinforcement learning agent as the basis for the decision-making in this iteration.

[0057] Step S2.8 involves the agent outputting the optimal threshold optimization action based on real-time state information, completing the online update of the threshold range. During execution, after receiving the real-time state vector, the reinforcement learning agent, with the core objective of maximizing cumulative reward, evaluates all executable actions in the action space through the built-in policy network, calculates the expected cumulative reward after each action is executed, and finally selects the optimal action with the highest expected reward. After outputting the optimal action, the agent updates the initial dynamic compliance threshold range online according to the action content, including fine-tuning the upper and lower limits of the threshold range, adjusting the corresponding feature weights, and optimizing the model regularization coefficients. The updated threshold range directly covers the original threshold range and serves as the new compliance judgment benchmark for the current process node. The entire decision-making and update process is completed automatically within the iteration cycle without manual intervention, achieving fully automatic online optimization of the threshold range.

[0058] Step S2.9 completes the closed-loop feedback and iterative convergence judgment to ensure that the agent's optimization continues to advance towards the preset goal. During execution, after the threshold interval is updated and put into use, the real-time control effect after the threshold interval update will be continuously collected, including new false alarm rate, false negative rate data, and the quality inspection results of the corresponding batch of finished products. According to the reward function set in step S2.6, the reward value corresponding to this action is calculated and fed back to the reinforcement learning agent. The agent will optimize its own policy network based on the feedback value to complete a complete iterative cycle. Subsequently, the above iterative cycle will be continuously executed. Each cycle will optimize the threshold interval and improve the agent's decision-making ability. After each iterative cycle, the control effect of the threshold interval will be judged. When the control effect reaches the preset stability standard, that is, the false alarm rate, false negative rate, and finished product qualification rate are all stable within the preset target range and there are no significant fluctuations for several consecutive iteration cycles, the iteration is judged to be converged, the cycle is stopped, and the final real-time dynamic compliance threshold interval adapted to the current production scenario is output. If the stability standard is not reached, the iterative cycle will continue to be executed to continuously optimize the threshold interval.

[0059] In one embodiment of the present invention, step S3 specifically includes the following sub-steps: S3.1 Match and compare the real-time operation data of each process traceability node with the real-time dynamic compliance threshold range of the corresponding node to determine whether the real-time operation data is within the threshold range. S3.2 Based on the magnitude of data deviation, duration of deviation, and independent marginal contribution of corresponding parameters, traceability nodes are divided into four compliance risk levels: no risk, low risk, medium risk, and high risk, and each node is marked with a corresponding level. S3.3 For traceability nodes marked as medium risk or above, they are determined to be abnormal traceability nodes. An abnormal traceability signal containing the abnormal node's identity, out-of-tolerance parameter information, risk level, and traceability time anchor point is generated and pushed to the visual traceability module.

[0060] Specifically, steps S3.1 to S3.3 in this embodiment are a complete process of online compliance determination, risk classification and abnormal triggering based on real-time dynamic compliance threshold range, which is a bridge connecting threshold control and full-process traceability.

[0061] Step S3.1 is to complete the online compliance determination based on the real-time dynamic compliance threshold range. During the execution process, for each traceability node of the entire production process, the operating data of the node is collected in real time. The real-time operating data is matched and compared with the current effective real-time dynamic compliance threshold range of the node to determine whether the real-time value of each parameter is within the corresponding threshold range. Parameters within the threshold range are determined to be compliant parameters. Parameters exceeding the threshold range are determined to be out-of-tolerance parameters, and the out-of-tolerance magnitude and duration are recorded.

[0062] Step S3.2 involves classifying and visualizing compliance risk levels. During execution, the risk level of traceability nodes is classified based on three core dimensions: the first is the magnitude of parameter deviation (the larger the deviation, the higher the risk level); the second is the duration of deviation (the longer the deviation lasts, the higher the risk level); and the third is the independent marginal contribution of the corresponding parameter, i.e., the parameter's weight in influencing the process quality result. Parameters with higher weights, even with smaller deviations, will correspond to a higher risk level. Based on a comprehensive assessment of these three dimensions, traceability nodes are classified into four compliance risk levels: no risk, low risk, medium risk, and high risk. No risk corresponds to nodes where all parameters are within the threshold range; low risk corresponds to nodes where non-core parameters have a small deviation over a very short period; medium risk corresponds to nodes where core parameters have a small deviation or non-core parameters have a large deviation; and high risk corresponds to nodes where core parameters have a large deviation or multiple parameters have a simultaneous deviation. After classifying the levels, each process node is visually labeled according to its corresponding level, providing a foundation for subsequent visualization.

[0063] The core of step S3.3 is to complete the identification of abnormal nodes and the generation and push of abnormal traceability signals. During the execution process, traceability nodes marked as medium risk or above are directly identified as abnormal traceability nodes. The parameter deviations of these nodes may affect the final quality of the product, requiring the initiation of a full-process traceability investigation. For each abnormal traceability node, a corresponding abnormal traceability signal is automatically generated. The signal contains the unique identifier of the abnormal node, detailed information on the out-of-tolerance parameters, the corresponding compliance risk level, the bound traceability time anchor point, and the corresponding production batch and product model information. After the abnormal traceability signal is generated, it is pushed to the visual traceability module in real time, simultaneously triggering the subsequent full-process root cause traceability process. At the same time, according to the risk level, the warning information is pushed to the corresponding production management personnel and quality inspection personnel to achieve timely response to the abnormality.

[0064] In one embodiment of the present invention, step S4, which completes the root cause localization of the entire process anomaly tracing link, specifically includes the following sub-steps: S4.1 Analyze the abnormal traceability signal, lock the production batch, process link and time window corresponding to the abnormal traceability node, and determine the boundary range of the whole process traceability; S4.2 In the distributed traceability database, retrieve the decoupled independent feature set, historical threshold range, real-time operation data, quality inspection data and operation record data of all process nodes within the locked traceability range; S4.3. Based on the independent marginal contribution of each parameter, quantify the impact weight of each process node on the abnormal result within the traceability range, locate the core root cause node and the corresponding parameter deviation, and reconstruct the entire abnormal impact chain from the root cause node to the abnormal node.

[0065] Specifically, steps S4.1 to S4.3 in this embodiment are a complete process for locating the root cause of anomalies throughout the entire process based on anomaly tracing signals. This process accurately locates the source of quality anomalies, reconstructs the entire chain propagation process of anomalies, and solves the pain points of traditional tracing methods, such as coarse granularity and inaccurate location.

[0066] Step S4.1 involves parsing the anomaly traceability signal and defining the boundary range of the entire process traceability. During execution, the anomaly traceability signal pushed in step S3 is first received. The signal is then parsed to extract core information, including the anomaly node's identity, traceability time anchor, production batch identifier, and risk level information. Based on the extracted information, the production batch corresponding to the anomaly traceability node is first identified, clarifying the product range that needs to be traced. Then, the entire process chain corresponding to this production batch is identified, clarifying all relevant process nodes from raw material entry to finished product inspection. Finally, based on the traceability time anchor, the time window of the anomaly occurrence is identified, clarifying the time range for data retrieval. Through these three dimensions of identification, the complete boundary range of the entire process traceability is ultimately determined, avoiding the low traceability efficiency caused by traversing all data, while ensuring that no processes and data related to the anomaly are missed.

[0067] Step S4.2 involves retrieving all relevant data within the traceability scope to provide complete data support for root cause localization. During execution, based on the traceability boundary range determined in step S4.1, relevant data for all process nodes within the traceability boundary range are retrieved from the distributed traceability database. Specifically, this includes the decoupled independent feature set of each process node, the historical threshold range effective during the production process of this batch, the real-time operating data of each node, the quality inspection data of the corresponding process, and the personnel operation record data during the production process. At the same time, the material batch information, equipment operating status information, and workshop environment data corresponding to this batch are also retrieved. After all data is retrieved, it is sorted and organized according to the sequence of processes and the time dimension to form a complete traceability dataset, ensuring that all possible influencing factors that may cause anomalies are covered during the root cause localization process.

[0068] Step S4.3 involves locating the core root cause and reconstructing the anomaly impact chain based on the traceability dataset. During execution, the independent marginal contribution of each parameter calculated in step S1 is used as the core weight to analyze the parameter data of all process nodes in the traceability dataset. The influence weight of each process node and each parameter within the traceability range on the final anomaly result is quantified. The higher the influence weight, the higher the probability that the node or parameter is the core cause of the anomaly. Through the quantification and sorting of weights, the core root cause node leading to the quality anomaly is finally located, as well as the corresponding parameter deviation in the core root cause node, clarifying the specific process, specific parameter, and specific time point where the anomaly occurred. After the root cause is located, based on the sequence of processes, the entire anomaly impact chain is reconstructed, from the parameter deviation of the core root cause node to the impact transmission of subsequent processes, and finally to the triggering of the warning by the anomaly node. This fully presents the entire life cycle of the anomaly's generation, transmission, and triggering, providing accurate basis for subsequent anomaly handling and quality optimization.

[0069] In one embodiment of the present invention, step S4 includes the following steps: [The text abruptly ends here, so the translation stops.] S4.4 Render a directed process topology diagram of the entire production process in the visualization interface, with a single process traceability node as the smallest visualization unit, and display the basic attribute information of each node accordingly. S4.5. Use floating color bands to render the real-time dynamic compliance threshold range of each process traceability node, and use gradient colors that correspond one-to-one with the compliance risk level to visually label each process node. S4.6 After receiving the abnormal traceability signal, the abnormal impact link and core root cause node are automatically highlighted in the process topology diagram, supporting users to drill down and trace back up from batch level to single parameter millisecond level traceability data. S4.7 Based on user interaction, the full-process traceability data, threshold change curves, risk level change records, quality inspection results and operation records within the corresponding traceability range are displayed synchronously on a unified timeline of the same visualization interface. It also includes the optimization steps for full-process traceability and closed-loop management: S4.8 Based on the core root cause nodes and parameter deviations of the location, a forward simulation model of the entire process parameters is constructed. The full process quality results after parameter adjustment are simulated to verify the accuracy of the root cause location. Based on the simulation verification results, S4.9 outputs parameter optimization suggestions for the corresponding process nodes. Simultaneously, it supplements the root cause data, optimization results, and quality feedback data of this anomaly tracing into the training sample set of the logistic regression model, completes the offline iterative update of the model, and forms a closed loop of full-process control for tracing, positioning, simulation, verification, optimization, and model update.

[0070] Specifically, steps S4.4 to S4.7 in this embodiment are the linked visualization display process of the whole process traceability data, and steps S4.8 to S4.9 are the closed-loop optimization process of the whole process traceability. The combination of the two parts realizes the intuitive presentation of traceability results and the continuous optimization of quality control, and completes the complete closed loop of this method from data to decision-making and from control to optimization.

[0071] The linked visualization display of steps S4.4 to S4.7 constructs a visual interactive system that is deeply integrated with the entire process of traceability and control, allowing operators in different roles to intuitively and quickly grasp the control status and anomaly traceability results of the entire production process.

[0072] Step S4.4 forms the foundation for building the visualization interface. During execution, a directed process topology diagram of the entire production process is rendered in the visualization interface. The directed process topology diagram is drawn entirely according to the process flow sequence of the actual production line. Each process traceability node is the smallest visualization unit, and each unit corresponds to a process node in the actual production line. The topology diagram displays the basic attribute information of the process node, including process name, equipment number, current operating status, and corresponding production batch information, allowing operators to intuitively grasp the overall layout and operating status of the production line.

[0073] Step S4.5 involves real-time visualization of threshold ranges and risk levels. During execution, for each process traceability node in the topology diagram, a floating color band is used to render the current dynamic compliance threshold range of that node. The upper and lower boundaries of the floating color band correspond to the upper and lower limits of the threshold range. Simultaneously, the current operating value of the parameter is marked in real-time within the color band, allowing operators to intuitively see the relationship between the parameter's operating position and the threshold range. Furthermore, a gradient color corresponding one-to-one with the compliance risk level is used to visually label each process node; for example, no risk is marked in green, low risk in blue, medium risk in yellow, and high risk in red. Operators can quickly identify risk nodes in the production line through color, without needing to check parameter values ​​one by one.

[0074] Step S4.6 involves highlighting the abnormal link and interactive drill-down of traceability data. During execution, when the visualization module receives an abnormal traceability signal, it automatically highlights the abnormal impact link and core root cause node in the process topology diagram. By using bold lines and flashing annotations, it intuitively presents the transmission path and source of the abnormality, allowing operators to locate the root cause of the abnormality at a glance. At the same time, the visualization interface supports bidirectional traceability data drill-down interaction. Operators can drill down from the batch-level production line overview to detailed data at the process level, parameter level, and millisecond level with one click. They can also trace back from the abnormal parameter node to the corresponding production batch, material batch, and the entire equipment operation process with one click, without switching between multiple interfaces, greatly improving the efficiency of traceability operations.

[0075] Step S4.7 is to achieve a unified timeline display of full-dimensional traceability data. During execution, based on the operator's interactive operations, a unified timeline is built in the same visualization interface. The full-process traceability data within the corresponding traceability scope, including parameter operation data of each process, threshold range change curves, risk level change records, quality inspection results, and personnel operation records, are all mapped onto this unified timeline. Operators can drag the timeline to view the synchronization status of all related processes and all related data at the same point in time, intuitively grasp the entire process evolution of abnormal events, quickly clarify the correlation between different parameters and different processes, and provide comprehensive decision support for abnormal handling.

[0076] The core of the full-process traceability closed-loop optimization in steps S4.8 to S4.9 is to break through the limitation of traditional traceability that can only be used for post-event accountability, and realize a complete closed loop from anomaly traceability to quality optimization, thereby promoting the continuous improvement of production control capabilities.

[0077] Step S4.8 verifies the accuracy of root cause localization through forward simulation, avoiding production risks caused by directly implementing optimized solutions. During execution, based on the core root cause nodes and corresponding parameter deviations identified in step S4.3, a forward simulation model of process parameters for the entire production process is constructed. This model is trained based on historical production data and can accurately simulate the quality results of the entire process under different parameter settings. The root cause parameter adjustment plan is input into the simulation model, and the quality results of the entire process after parameter adjustment are simulated. This allows us to check whether the adjusted parameters can resolve the current quality anomalies and whether they will cause other new quality risks. The accuracy of root cause localization is verified through the simulation results, ensuring that the root cause localization is unbiased and the optimization plan is effective.

[0078] Step S4.9 involves outputting the optimization plan and completing the iterative update of the model, forming a closed-loop control system for the entire process. During execution, based on the simulation verification results, actionable parameter optimization suggestions for the corresponding process nodes are output and provided to production management personnel for on-site adjustments to resolve current quality anomalies. Simultaneously, the entire process data of this anomaly tracing, including root cause node data, parameter deviation data, optimization plan, and optimized quality feedback data, is added to the training sample set of the logistic regression model for offline iterative updates. This allows the model to learn the characteristics of this anomaly event, further improving the accuracy of subsequent threshold modeling and anomaly warning. Through this process, a closed-loop control system for anomaly tracing, root cause location, simulation verification, parameter optimization, and model updates is ultimately formed, achieving continuous iterative improvement in production quality control capabilities and completely solving the industry pain point of recurring and intractable quality problems in traditional traceability technologies.

[0079] By constructing a correlation model between parameters and quality results using decoupled, clean feature data, and utilizing a logistic regression model to generate an initial dynamic compliance threshold range adapted to the current production scenario, this approach effectively replaces traditional fixed thresholds, improving the accuracy of threshold generation and scenario adaptability from the modeling source. This method abandons the traditional extensive approach of relying on manual experience and single statistical values ​​to set thresholds. Instead, it uses the quality of the process as the learning target, fully learning the inherent patterns between parameters and quality in historical data. Combined with multi-dimensional scenario information such as current product model, material batch, equipment status, and environmental conditions, it outputs a dynamic threshold range with statistical confidence. This makes the threshold no longer a static fixed value, but an adaptive benchmark strongly correlated with the production scenario. This solves the long-standing problems of poor adaptability, disconnect from actual working conditions, susceptibility to working condition drift, and frequent false alarms and missed alarms associated with traditional fixed thresholds. This makes the quality control benchmark more scientific, reasonable, and highly aligned with the actual production status on site.

[0080] By introducing a reinforcement learning agent to perform online self-iterative optimization of the initial dynamic compliance threshold range, and using real-time operating conditions, threshold control effectiveness, and finished product quality as closed-loop feedback, a set of intelligent threshold control mechanisms that can learn autonomously and continuously evolve is constructed. The agent optimizes the false alarm rate, false negative rate, and finished product qualification rate. Under the premise of meeting industrial physical boundary constraints, it automatically completes the fine-tuning of the upper and lower limits of the threshold, updates the feature weights, and optimizes the model regularization coefficients. This enables the threshold range to continuously and adaptively correct itself as the production environment fluctuates, material batches change, equipment performance degrades, and process parameter drifts, without the need for frequent manual intervention and calibration. This mechanism upgrades the system from a one-time static modeling to a dynamic control mode of long-term online iterative evolution, further reducing the false alarm rate and false negative rate of early warning, improving the accuracy of anomaly identification and the long-term stability of the system, and significantly improving the level of intelligent quality control throughout the entire production process.

[0081] A full-process visual traceability system based on production site data includes multi-source heterogeneous data acquisition, anomaly warning, and visual traceability of the entire production process. The system is as follows: For the strongly coupled operating condition parameter data of each process node in the entire production process, a multivariate adaptive regression model is used to decouple the parameters, remove the cross-coupling influence between multiple parameters, quantify the independent marginal contribution of each single parameter to the quality result of the corresponding process, and generate and output the decoupled independent feature set of each process traceability node. An adaptive modeling framework integrating reinforcement learning and logistic regression is constructed. Historical operating data, material batch data, and quality inspection data corresponding to each traceability node are used as training aids. First, the mapping relationship between decoupled independent features and process quality results is fitted by a logistic regression model to generate the initial dynamic compliance threshold range for each traceability node. Then, through a pre-set reinforcement learning agent, the initial dynamic compliance threshold range is iteratively optimized online with real-time operating conditions, threshold control effects, and finished product quality results as closed-loop feedback, and the real-time dynamic compliance threshold range adapted to the current production scenario is output. Online compliance assessment is performed on the real-time operational data of each process traceability node, abnormal traceability nodes are identified and their corresponding compliance risk levels are marked, and abnormal traceability signals are generated simultaneously. Complete the root cause location of the entire process of anomaly tracing, and simultaneously realize the linked visualization display of the entire production process traceability data, real-time dynamic compliance threshold range, compliance risk level and anomaly tracing link on the visual interface.

[0082] The embodiments of this example have been described above. However, this example is 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 many other forms based on the guidance of this example, and all of them are within the protection scope of this example.

Claims

1. A full-process visual traceability method based on production site data, including multi-source heterogeneous data collection and anomaly early warning and visual traceability links throughout the entire production process, characterized in that... The steps are as follows: S1. Decouple the strongly coupled operating condition parameter data of each process node in the entire production process based on the multivariate adaptive regression model, remove the cross-coupling influence between multiple parameters, quantify the independent marginal contribution of each single parameter to the quality result of the corresponding process, and generate and output the decoupled independent feature set of each process traceability node. S2. Based on the decoupled independent feature set in step S1, construct an adaptive modeling framework that integrates reinforcement learning and logistic regression. Use historical working condition data, material batch data and quality inspection data corresponding to each traceability node as training aids. First, fit the mapping relationship between the decoupled independent features and the process quality results through the logistic regression model to generate the initial dynamic compliance threshold range for each traceability node. Then, through a pre-set reinforcement learning agent, with real-time working conditions, threshold control effects, and finished product quality results as closed-loop feedback, the initial dynamic compliance threshold range is optimized online through self-iteration, and a real-time dynamic compliance threshold range adapted to the current production scenario is output. S3. Based on the real-time dynamic compliance threshold range of step S2, online compliance judgment is performed on the real-time operation data of each process traceability node, abnormal traceability nodes are identified and the corresponding compliance risk level is marked, and abnormal traceability signals are generated synchronously. S4. Based on the abnormal traceability signal in step S3, complete the root cause location of the entire abnormal traceability chain, and simultaneously realize the linkage visualization display of the entire production process traceability data, real-time dynamic compliance threshold range, compliance risk level and abnormal traceability chain on the visualization interface.

2. The method for full-process visualization and traceability based on production site data according to claim 1, characterized in that, Step S1 includes the following steps: S1.

1. For the collected strongly coupled working condition parameter data of each process node, align the time dimension with the production cycle time of the corresponding process, remove invalid null values ​​and abnormal jump data, and generate a standardized coupled parameter dataset. S1.2 The quality inspection result of the current process is the output dependent variable, and the parameters of each working condition in the standardized coupled parameter dataset are the input independent variables to construct a multivariate adaptive regression equation with an adaptive regularization term. S1.3 Solve the partial derivatives of the multivariate adaptive regression equation, remove the cross-coupling terms between the input independent variables, calculate the pure influence coefficient of each input independent variable on the output dependent variable, and quantify the independent marginal contribution of the corresponding single parameter to the process quality result based on the pure influence coefficient. S1.4 Sort each single parameter based on independent marginal contribution, remove redundant parameters with contribution below the preset threshold, retain effective core parameters, and generate decoupled independent feature sets for the corresponding process traceability nodes.

3. The method for full-process visual traceability based on production site data according to claim 2, characterized in that, Step S1 further includes the following sub-steps: S1.5 Decouple each single-parameter feature in the independent feature set and bind a unique, tamper-proof, traceable time anchor point, process identity identifier, and production batch identifier that are strongly associated with the corresponding process production cycle time. S1.6 Perform feature consistency and integrity checks on the decoupled independent feature sets of the bound traceability anchors. After passing the checks, store them in the distributed traceability database for subsequent threshold modeling and full-process traceability calls.

4. The method for full-process visualization and traceability based on production site data according to claim 1, characterized in that, In step S2, the initial dynamic compliance threshold range for each traceability node is generated using a logistic regression model, including the following steps: S2.

1. The historical decoupled independent feature set of the corresponding process traceability node is used as the input sample. The process quality qualified / unqualified result of the corresponding historical batch is used as the binary label. The corresponding historical working condition data, material batch data, and equipment operation data are matched to construct a standardized training sample set for the logistic regression model. S2.

2. Supervised training of the logistic regression model is performed based on the standardized training sample set. The nonlinear mapping relationship between each single parameter feature in the decoupled independent feature set and the process quality result is fitted to complete the model convergence verification and generalization accuracy verification. S2.3 Based on the preset confidence level, and combined with the product model, material batch, equipment status, and environmental parameters of the current production scenario, the trained logistic regression model outputs the upper and lower limits of the initial dynamic compliance threshold range for the corresponding traceability node.

5. The method for full-process visual traceability based on production site data according to claim 4, characterized in that, In step S2, the construction of the preset reinforcement learning agent specifically includes the following sub-steps: S2.

4. Using the real-time working conditions of the entire production process as the intelligent agent's interaction environment, a multi-dimensional state space is constructed. The state vector of the state space includes the decoupled independent feature set of the current process, the initial dynamic compliance threshold range, real-time working condition parameters, real-time quality inspection results, equipment operating status, and compliance risk level. S2.5 Construct the executable action space of the intelligent agent. The action space includes fine-tuning actions of the upper and lower limits of the initial dynamic compliance threshold range, adjustment actions of the weights of each feature in the decoupled independent feature set, and regularization coefficient optimization actions of the logistic regression model. Simultaneously set the physical boundary constraints and adjustment step size rules for each action. S2.6 The core optimization targets for threshold control are false alarm rate, false negative rate, and finished product qualification rate. A reward function is set: when the agent performs an action, if the false alarm rate and false negative rate of threshold control decrease and the finished product qualification rate increases, a positive reward is given; if the false alarm rate and false negative rate increase or the finished product qualification rate decreases, a negative penalty is given.

6. The method for full-process visualization and traceability based on production site data according to claim 5, characterized in that, In step S2, the initial dynamic compliance threshold range is optimized online through a reinforcement learning agent, specifically including the following sub-steps: S2.

7. Collect the real-time state vector of the current production scenario according to the preset iteration cycle, and input it into the reinforcement learning agent after standardization processing. S2.8 The agent, based on the current state vector, aims to maximize the cumulative reward and outputs the optimal action through the policy network to complete the online update of the initial dynamic compliance threshold range. S2.

9. Collect the real-time control effect and finished product quality results after the threshold update, calculate the corresponding reward value and feed it back to the agent to complete one iteration cycle; continue to execute the iteration cycle until the control effect of the threshold range reaches the preset stable standard, and output the final real-time dynamic compliance threshold range adapted to the current scenario.

7. The method for full-process visual traceability based on production site data according to claim 1, characterized in that, Step S3 specifically includes the following sub-steps: S3.1 Match and compare the real-time operation data of each process traceability node with the real-time dynamic compliance threshold range of the corresponding node to determine whether the real-time operation data is within the threshold range. S3.2 Based on the magnitude of data deviation, duration of deviation, and independent marginal contribution of corresponding parameters, traceability nodes are divided into four compliance risk levels: no risk, low risk, medium risk, and high risk, and each node is marked with a corresponding level. S3.3 For traceability nodes marked as medium risk or above, they are determined to be abnormal traceability nodes. An abnormal traceability signal containing the abnormal node's identity, out-of-tolerance parameter information, risk level, and traceability time anchor point is generated and pushed to the visual traceability module.

8. The method for full-process visualization and traceability based on production site data according to claim 1, characterized in that, In step S4, the root cause localization of the entire process anomaly tracing chain is completed, which specifically includes the following sub-steps: S4.1 Analyze the abnormal traceability signal, lock the production batch, process link and time window corresponding to the abnormal traceability node, and determine the boundary range of the whole process traceability; S4.2 In the distributed traceability database, retrieve the decoupled independent feature set, historical threshold range, real-time operation data, quality inspection data and operation record data of all process nodes within the locked traceability range; S4.

3. Based on the independent marginal contribution of each parameter, quantify the impact weight of each process node on the abnormal result within the traceability range, locate the core root cause node and the corresponding parameter deviation, and reconstruct the entire abnormal impact chain from the root cause node to the abnormal node.

9. The method for full-process visual traceability based on production site data according to claim 8, characterized in that, Step S4 includes the following steps of linked visualization display: S4.4 Render a directed process topology diagram of the entire production process in the visualization interface, with a single process traceability node as the smallest visualization unit, and display the basic attribute information of each node accordingly. S4.

5. Use floating color bands to render the real-time dynamic compliance threshold range of each process traceability node, and use gradient colors that correspond one-to-one with the compliance risk level to visually label each process node. S4.6 After receiving the abnormal traceability signal, the abnormal impact link and core root cause node are automatically highlighted in the process topology diagram, supporting users to drill down and trace back up from batch level to single parameter millisecond level traceability data. S4.7 Based on user interaction, the full-process traceability data, threshold change curves, risk level change records, quality inspection results and operation records within the corresponding traceability range are displayed synchronously on a unified timeline of the same visualization interface. It also includes the optimization steps for full-process traceability and closed-loop management: S4.8 Based on the core root cause nodes and parameter deviations of the location, a forward simulation model of the entire process parameters is constructed. The full process quality results after parameter adjustment are simulated to verify the accuracy of the root cause location. Based on the simulation verification results, S4.9 outputs parameter optimization suggestions for the corresponding process nodes. Simultaneously, it supplements the root cause data, optimization results, and quality feedback data of this anomaly tracing into the training sample set of the logistic regression model, completes the offline iterative update of the model, and forms a closed loop of full-process control for tracing, positioning, simulation, verification, optimization, and model update.

10. A full-process visual traceability system based on production site data, characterized in that: This system is used to execute the method as described in any one of claims 1 to 9, including the acquisition of multi-source heterogeneous data throughout the entire production process and the anomaly early warning and visual traceability stage, characterized in that the system is as follows: For the strongly coupled operating condition parameter data of each process node in the entire production process, a multivariate adaptive regression model is used to decouple the parameters, remove the cross-coupling influence between multiple parameters, quantify the independent marginal contribution of each single parameter to the quality result of the corresponding process, and generate and output the decoupled independent feature set of each process traceability node. An adaptive modeling framework integrating reinforcement learning and logistic regression is constructed. Historical operating data, material batch data, and quality inspection data corresponding to each traceability node are used as training aids. First, the mapping relationship between decoupled independent features and process quality results is fitted by a logistic regression model to generate the initial dynamic compliance threshold range for each traceability node. Then, through a pre-set reinforcement learning agent, the initial dynamic compliance threshold range is iteratively optimized online with real-time operating conditions, threshold control effects, and finished product quality results as closed-loop feedback, and the real-time dynamic compliance threshold range adapted to the current production scenario is output. Online compliance assessment is performed on the real-time operational data of each process traceability node, abnormal traceability nodes are identified and their corresponding compliance risk levels are marked, and abnormal traceability signals are generated simultaneously. Complete the root cause location of the entire process of anomaly tracing, and simultaneously realize the linked visualization display of the entire production process traceability data, real-time dynamic compliance threshold range, compliance risk level and anomaly tracing link on the visual interface.

Images