A full-process monitoring and data management system for an inward-opening window assembly production line

Abstract

This invention discloses a full-process monitoring data management system for an inward-opening window assembly line, belonging to the field of industrial automation and data processing technology. It includes a distributed data acquisition and preprocessing module, used to collect sensor data from various production nodes on the production line and attach high-precision timestamps and material supplier identifiers, providing traceable foundational data for subsequent latency calculations and semantic consistency analysis. A data semantic consistency quantification module calculates the KL divergence value of the current sensor data distribution relative to a preset reference data distribution model and uses an exponential function for mapping, generating a quantified semantic consistency health index. This enables accurate assessment of data quality changes caused by the mixing of materials from new and old suppliers. This invention provides a solid and reliable input for subsequent model evaluation and decision-making.

Description

Technical Field

[0001] This invention relates to the field of industrial automation and data processing, specifically to a full-process monitoring and data management system for an inward-opening window assembly production line. Background Technology

[0002] In modern manufacturing, especially for companies with globally distributed supply chains, the stability of production lines and the uniformity of product quality are crucial. Existing production line monitoring systems typically rely on centralized data collection and analysis, aiming to improve production efficiency by predicting production defects through big data and machine learning models. However, in cross-regional, multi-node production networks, such systems face two major challenges: first, data latency, where production data from different factories arrives at the central processing system at inconsistent times due to network environment, geographical location, and other factors, preventing the global model from making decisions based on real-time and complete status; second, semantic consistency issues, where when new and old materials from different suppliers are mixed during production, even sensor data for the same process may have altered physical or chemical properties, causing data distribution drift. This semantic inconsistency severely contaminates training data, leading to a sharp decline in the accuracy of predictive models and even triggering false warnings. Current solutions often employ static data cleaning strategies or simple delay mechanisms, failing to dynamically adapt to changes in data quality and failing to effectively balance ensuring the stability of the global model with maintaining the normal operation of local factories.

[0003] The information disclosed in the background section above is only intended to enhance the understanding of the background of this disclosure, and therefore may include information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] The purpose of this invention is to provide a full-process monitoring and data management system for an inward-opening window assembly production line to solve the problems mentioned in the background art.

[0005] The technical solution of the present invention includes:

[0006] The distributed data acquisition and preprocessing module is used to collect sensor data on the production line at various production nodes and attach high-precision timestamps and material supplier identifiers, providing traceable basic data for subsequent latency calculation and semantic consistency analysis.

[0007] The data semantic consistency quantification module calculates the KL divergence value of the current sensor data distribution relative to the preset reference data distribution model, and uses an exponential function for mapping to generate a quantified semantic consistency health index, so as to achieve accurate assessment of data quality changes caused by the mixing of materials from new and old suppliers.

[0008] The Federal Defect Prediction and Model Trust Assessment module combines the semantic consistency health index with the defect prediction recall rate of the global defect prediction model, and generates a comprehensive model trust index through weighted geometric average calculation. This index is used to simultaneously reflect the dual guarantee of predictive ability and data quality.

[0009] The dynamic balancing and strategic decision-making module calculates the safe distance between the current prediction accuracy and the preset failure boundary, and calculates the distrust level by combining the model trust index, thereby generating a strategic adjustment factor to quantify system risk. The strategic adjustment factor is compared with a preset threshold to generate an isolation signal to trigger the isolation mechanism or a normal operation signal to maintain the current state. In this way, while ensuring the basic operation of local production nodes, it adaptively protects the stability of the global model and the long-term interests of the entire production network.

[0010] Preferably, the generation process of the semantic consistency health index is as follows:

[0011] Obtain a reference data distribution model based on historical material data from a high-quality single supplier, and obtain the current empirical probability distribution based on current sensor data;

[0012] Calculate the KL divergence value of the current empirical probability distribution relative to the reference data distribution model to quantify the degree of deviation of the current data distribution;

[0013] By mapping the KL divergence value using an exponential function with a preset decay coefficient, a semantic consistency health index is generated, thereby converting the unbounded divergence value into an intuitive and standardized data quality score.

[0014] Preferably, it also includes an adaptive data synchronization and latency compensation module, wherein the decision processing process of the adaptive data synchronization and latency compensation module is as follows:

[0015] The semantic consistency health index is compared with the confidence threshold determined through offline simulation experiments;

[0016] When the semantic consistency health index is higher than the confidence threshold, the data source quality is determined to be stable and a prediction filling signal is generated. The pre-trained autoregressive integral moving average model is then started to fill the data, so as to significantly reduce the global data synchronization latency with acceptable accuracy loss.

[0017] When the semantic consistency health index is lower than or equal to the confidence threshold, the data distribution is determined to be unreliable and a strict wait signal is generated, suspending any data filling operation to ensure the authenticity of the original data and avoid introducing uncontrollable noise to pollute the global model.

[0018] Preferably, the generation process of the model trust index is as follows:

[0019] Obtain the defect prediction recall rate of the global defect prediction model at the current production node;

[0020] By combining the weighting coefficients preset by managers based on the strategic goals of the enterprise at different stages, the defect prediction recall rate and semantic consistency health index are calculated by weighted geometric average to generate a comprehensive model trust index, ensuring that the index is highly sensitive to any decline in predictive ability or data quality.

[0021] Preferably, the generation process of the strategic adjustment factor is as follows:

[0022] Obtain the current prediction accuracy of the plant and the preset failure boundary set by the production process experts, and calculate the difference between the two to determine the safe distance representing the system's buffer space;

[0023] The model's trust index is reverse-processed to obtain the distrust index;

[0024] Dividing the level of distrust by the safe distance generates a strategic adjustment factor, thereby using the buffer space of the system's distance from failure as a risk amplifier, so that the closer the prediction accuracy is to the failure boundary, the more significant the increase in risk assessment results.

[0025] Preferably, the final decision and response of the dynamic equilibrium and strategic decision-making module are as follows:

[0026] When the strategic adjustment factor exceeds the preset adjustment factor threshold, an isolation signal is generated, automatically disconnecting the corresponding production node from the federated learning network and instructing it to degrade to local statistical process control mode, sacrificing the short-term advanced prediction function of the node to protect the purity of the global model.

[0027] When the strategic adjustment factor does not exceed the preset adjustment factor threshold, a normal operation signal is generated to maintain the normal training and prediction status of the production node in the federated learning network.

[0028] This invention provides an improved data management system for the entire process monitoring of an inward-opening window assembly production line, which has the following improvements and advantages compared with the prior art:

[0029] 1. This system transforms abstract data quality issues into precise and measurable technical indicators. Existing technologies struggle to effectively assess the impact of mixed materials from multiple suppliers on production data consistency. This system, by constructing a data semantic consistency quantification module, can calculate the statistical deviation between the current sensor data distribution and a preset reference data distribution model, and map it into a standardized semantic consistency health index. This transforms the vague concept of data quality into a quantitative basis that can accurately and sensitively capture subtle shifts in data distribution caused by material changes, providing a solid and reliable input for subsequent model evaluation and decision-making.

[0030] 2. This system constructs a model reliability evaluation system that can simultaneously reflect both predictive capability and data quality. Traditional model evaluation methods usually only focus on the accuracy or recall of defect predictions, ignoring the potential risks brought about by fluctuations in the quality of input data. This system innovatively sets up a federated defect prediction and model trust evaluation module, which uses a weighted geometric average of defect prediction recall and semantic consistency health index to generate a comprehensive model trust index. The inherent logic of this index ensures that it is highly sensitive to any decline in predictive performance or data quality, thereby enabling a more comprehensive and profound insight into the reliability of the prediction behavior itself and avoiding the risk of drawing wrong conclusions based on unreliable data.

[0031] 3. This system achieves forward-looking quantification and non-linear amplification of global failure risk. Unlike existing passive response management models, this system establishes a dynamic balancing and strategic decision-making module to proactively assess system risk. This module calculates the safe distance between the current prediction accuracy and the preset failure boundary, and combines this with the inverse indicator of model confidence—distrust—to generate a strategic adjustment factor. This design cleverly uses the safe distance as a risk amplifier, causing the risk assessment result to increase sharply as the system's performance approaches the failure boundary. This non-linear risk quantification mechanism can provide earlier and more significant early warnings of systemic risks than traditional linear monitoring methods.

[0032] 4. Based on forward-looking risk quantification, this system establishes an adaptive global network protection mechanism. In a distributed production network, the failure of a single node may pollute the global model and cause a chain reaction. By comparing strategic adjustment factors with preset thresholds, this system can autonomously generate isolation signals or normal operation signals. When the risk of a production node is determined to be too high, the system will automatically disconnect it from the federated learning network and downgrade it to local statistical process control mode. This dynamic balance strategy, which sacrifices the short-term advanced functions of local nodes in exchange for the purity and stability of the global model, solves the risk diffusion problem in distributed production networks and ensures the long-term healthy and efficient operation of the entire production network. Attached Figure Description

[0033] The present invention will be further explained below with reference to the accompanying drawings and embodiments:

[0034] Figure 1 This is a flowchart of the system of the present invention. Detailed Implementation

[0035] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments. Example 1

[0036] Please see Figure 1 This invention provides a full-process monitoring and data management system for an inward-opening window assembly production line, comprising:

[0037] The distributed data acquisition and preprocessing module is used to collect sensor data on the production line at various production nodes and attach high-precision timestamps and material supplier identifiers, providing traceable basic data for subsequent latency calculation and semantic consistency analysis.

[0038] The data semantic consistency quantification module calculates the KL divergence value of the current sensor data distribution relative to the preset reference data distribution model, and uses an exponential function for mapping to generate a quantified semantic consistency health index, so as to achieve accurate assessment of data quality changes caused by the mixing of materials from new and old suppliers.

[0039] The Federal Defect Prediction and Model Trust Assessment module combines the semantic consistency health index with the defect prediction recall rate of the global defect prediction model, and generates a comprehensive model trust index through weighted geometric average calculation. This index is used to simultaneously reflect the dual guarantee of predictive ability and data quality.

[0040] The dynamic balancing and strategic decision-making module calculates the safe distance between the current prediction accuracy and the preset failure boundary, and calculates the distrust level by combining the model trust index, thereby generating a strategic adjustment factor to quantify system risk. The strategic adjustment factor is compared with a preset threshold to generate an isolation signal to trigger the isolation mechanism or a normal operation signal to maintain the current state, thereby adaptively protecting the stability of the global model and the long-term interests of the entire production network while ensuring the basic operation of local production nodes.

[0041] A data management system for monitoring the entire process of an inward-opening window assembly line has been disclosed. The system aims to address the challenges of predictive model failure caused by semantic inconsistencies due to data latency and material mixing in cross-regional, multi-supplier production networks. Through the collaborative work of various functional modules, the system achieves quantitative assessment of data quality, dynamic calculation of model trust, and adaptive strategic decision-making based on system risks, thereby maximizing the long-term production benefits of the group as a whole while ensuring the stable operation of local production nodes.

[0042] In this embodiment, the system includes four core functional modules;

[0043] A distributed data acquisition and preprocessing module is deployed at each production node. The purpose of this module is to acquire raw data from the sensor array on the production line and standardize it, providing a traceable foundation for all subsequent analyses. In this embodiment, this module is configured to independently acquire sensor data, such as pressure, temperature, and flow rate, from key processes like gluing, corner assembly, and pressing on the internal window assembly line at each production node in real time. Each acquired data record is automatically appended with a high-precision local timestamp and a source identifier. The timestamp refers to the precise local time when the data was acquired, serving as a benchmark for subsequent latency calculations; its source is the clock synchronization signal of the local server on the production node. The source identifier is a composite code containing the factory ID, production line ID, and the batch number of the current material supplier, ensuring that each piece of data can be traced back to its physical source, which is the work order and material information from the production execution system. This operation provides a traceable foundation for subsequent latency calculations and semantic consistency analysis.

[0044] The data semantic consistency quantification module is configured to transform abstract data quality issues into measurable mathematical indicators. The purpose of this module is to assess the degree of change in data distribution caused by changing material suppliers. In this embodiment, the module calculates the statistical difference between the current sensor data distribution and a preset reference data distribution model, i.e., the KL divergence value, and uses an exponential function for normalization mapping to generate a quantified semantic consistency health index. This index accurately reflects the changes in data quality caused by the mixing of materials from old and new suppliers.

[0045] The Federated Defect Prediction and Model Trust Evaluation module is configured to build a global prediction model while protecting data privacy and continuously evaluate its reliability. The module's purpose is not only to predict production defects but also to assess the reliability of the prediction itself. In this embodiment, the module runs within a federated learning framework, where each node only uploads model parameters, not raw data, to aggregate and train the global defect prediction model. Simultaneously, the module couples the aforementioned semantic consistency health index with the defect prediction recall rate of the global model at each node, calculating and generating a comprehensive model trust index using a weighted geometric average algorithm. This index simultaneously reflects both the predictive power of the prediction model and the quality of the data foundation.

[0046] The dynamic balancing and strategic decision-making module serves as the core decision-making unit of the system. Its purpose is to make optimal adaptive adjustments based on a quantitative assessment of the overall system risk, achieving a dynamic balance between local stability and global optimality. In this embodiment, the module continuously monitors the safe distance between the prediction accuracy of each production node and its preset failure boundary. This safe distance is combined with a model confidence index, and a strategic adjustment factor is calculated using a risk amplification model to quantify the current risks faced by the system. The module compares this real-time calculated strategic adjustment factor with a preset decision threshold, and the comparison result generates a decision signal. This decision signal is used to trigger an isolation mechanism, temporarily disconnecting high-risk nodes or maintaining the normal operation of the current node.

[0047] The system provided in this embodiment, through the organic combination of the above modules, constructs a complete closed-loop management system from data source quality control to model application risk assessment, and then to system-level strategic adjustments; it solves the problem that static data processing strategies in the prior art cannot cope with the dynamic changes in data latency and semantic inconsistency in distributed production networks; the system can adaptively protect the stability of the global model and the long-term interests of the entire production network, while maximizing the continuous operation of local production nodes, and achieving long-term maximization of production efficiency in the context of complex supply chains.

[0048] The process of generating the semantic consistency health metric is as follows:

[0049] Obtain a reference data distribution model based on historical material data from a high-quality single supplier, and obtain the current empirical probability distribution based on current sensor data;

[0050] Calculate the KL divergence value of the current empirical probability distribution relative to the reference data distribution model to quantify the degree of deviation of the current data distribution;

[0051] By mapping the KL divergence value using an exponential function with a preset decay coefficient, a semantic consistency health index is generated, thereby converting the unbounded divergence value into an intuitive and standardized data quality score.

[0052] This embodiment is a specific implementation of the data semantic consistency quantification module described above. Its core lies in accurately quantifying the abstract concept of data quality using mathematics.

[0053] The generation process of the semantic consistency health index in this module is limited; the calculation process is as follows: the system obtains a reference data distribution model; the reference data distribution model refers to a probability distribution model that represents the sensor data behavior pattern of a specific process under ideal production conditions. Its role is to provide a gold standard or benchmark for evaluating the current data quality. It is constructed by using non-parametric statistical methods such as kernel density estimation based on sensor data generated from materials from a single supplier that have been used for a long time in history and have been verified to be of high quality and stable performance.

[0054] High quality and stable performance refer to products manufactured from materials that have a defect rate below a preset threshold and exhibit minimal historical batch fluctuations in key sensor data statistical characteristics, such as mean and variance, after rigorous quality control verification in long-term production practice. When constructing a reference data distribution model, non-parametric statistical methods such as Gaussian mixture model kernel density estimation can be used. Taking KDE as an example, its core lies in selecting a suitable kernel function, such as a Gaussian kernel, and a bandwidth parameter. The bandwidth parameter can be optimized through cross-validation or rule-based methods, such as Silvermans' Rule of Thumb, to balance the model's smoothness with its ability to capture data details. The construction process can be represented as:

[0055] in, For the estimated probability density function, For high-quality data sample size, For kernel function, For bandwidth parameters; : data points; A single data sample, representing a single data point from a high-quality dataset;

[0056] At the same time, the system acquires the current empirical probability distribution based on the data collected in real time by the current sensors;

[0057] The current empirical probability distribution can be constructed using either a sliding time window or a fixed sample batch method. In real-time monitoring, a sliding time window strategy is recommended, for example, statistically analyzing all data points collected within the last 60 seconds every second to construct the current empirical probability distribution. This method ensures that the distributed model reflects the latest production status in real time and smoothly captures dynamic changes in data distribution.

[0058] Based on the above model and distribution, the module calculates the KL divergence value of the current empirical probability distribution relative to the reference data distribution model. The KL divergence value is an asymmetric measure used to measure the difference between two probability distributions. Its function is to quantitatively describe the degree to which the current data distribution deviates from the ideal state due to factors such as material changes. It comes from the standard calculation method in information theory.

[0059] The module uses an exponential function with a preset decay coefficient to map the KL divergence value to generate the final semantic consistency health index; the calculation formula is designed as follows:

[0060] in, The semantic consistency health index is a standardized score within the interval (0, 1), with a data type of floating-point number, and is calculated by this formula; Euler's number is the base of the natural logarithm. : KL divergence value, used to measure the difference between two distributions, data type is non-negative floating point, calculated by the previous steps; The current empirical probability distribution is a statistical description of the current batch of material data, constructed by this module based on real-time sensor data. Reference data distribution model: This represents the data distribution under ideal conditions and is a preset model in the system. The attenuation coefficient is a dimensionless, preset parameter used to adjust the sensitivity of the health score to changes in data distribution. To clarify the calibration process of this attenuation coefficient, it is not a runtime variable but is predetermined through analysis of the calibration dataset. Specifically, it is calibrated through regression analysis of historical data. A specific case in history where a significant increase in production defect rate due to material changes is selected, and the KL divergence value between the sensor data distribution and the reference data distribution model in that case is extracted and denoted as the calibration divergence. Simultaneously, a warning health level is set, denoted as the standard health level. For example, 0.7; by solving the equation Inverse solution The value of ;

[0061] To ensure the attenuation coefficient To determine the validity of the calibration, the calibration process should include the following steps:

[0062] Data Collection: Several typical cases in history where production defect rates increased significantly due to material changes were selected. Sensor data for each case were extracted, and the KL divergence value relative to the reference distribution model was calculated to form a set of divergence values. ;

[0063] Setting Health Levels: Collaborate with the production quality department to set health level thresholds corresponding to warning states, for example... This value represents the critical data quality level that the system deems requires close attention.

[0064] Solution optimization: Set the health threshold and the mean or median of set D Substitute into the formula By solving the equation in reverse, we can obtain... The optimal value:

[0065] in, Attenuation coefficient; Natural logarithm; : Define the health level, which is the health level threshold corresponding to the warning state, for example, 0.7; : Calibration divergence, the KL divergence value obtained by analyzing specific cases in history where the production defect rate increased significantly due to material changes;

[0066] This method ensures It can be calibrated based on actual production history data, making the health rating system business-relevant;

[0067] This step transforms the unbounded divergence value into an intuitive and standardized data quality score;

[0068] Through the specific implementation methods described above, this invention transforms the vague concept of data quality into a precise, continuous, and standardized health indicator. Compared to traditional rule-based static cleaning strategies, this method can more sensitively and accurately capture subtle data distribution shifts caused by material mixing, providing reliable and quantifiable input for subsequent adaptive data synchronization, model trust assessment, and strategic decision-making.

[0069] Example 2

[0070] It also includes an adaptive data synchronization and latency compensation module. The decision-making process of the adaptive data synchronization and latency compensation module is as follows:

[0071] The semantic consistency health index is compared with the confidence threshold determined through offline simulation experiments;

[0072] When the semantic consistency health index is higher than the confidence threshold, the data source quality is determined to be stable and a prediction filling signal is generated. The pre-trained autoregressive integral moving average model is then started to fill the data, so as to significantly reduce the global data synchronization latency with acceptable accuracy loss.

[0073] When the semantic consistency health index is lower than or equal to the confidence threshold, the data distribution is determined to be unreliable and a strict wait signal is generated, suspending any data filling operation to ensure the authenticity of the original data and avoid introducing uncontrollable noise to pollute the global model.

[0074] This embodiment adds an adaptive data synchronization and latency compensation module to the above system; the purpose of this module is to handle the problem of asynchronous arrival of data from different production nodes while taking into account the timeliness and authenticity of the data.

[0075] The decision-making process of the adaptive data synchronization and latency compensation module is as follows:

[0076] This module calculates the real-time semantic consistency health index from the data semantic consistency quantification module. With dynamically determined confidence thresholds Compare; confidence threshold This refers to the critical point used to determine whether the data quality is reliable enough to support predictive imputation; its function is to serve as a decision-making basis for switching data processing strategies. The working principle and setting method of this threshold are as follows: through a series of offline simulation experiments, it systematically evaluates different... At the horizontal level, the impact of the predictive filling strategy on the accuracy and stability of the global defect prediction model; Set as the critical point, when Above this point, the incremental model prediction error introduced by data filling is lower than the preset, acceptable error tolerance. Conversely, it will exceed this tolerance level; this error tolerance level The source is the specific quality control indicators formulated by the production quality management department based on long-term benefit assessments;

[0077] When in real time The indicator is above the confidence threshold When the system determines that the data source arriving late is of stable quality, the module generates a prediction fill signal. The prediction fill signal is an internal instruction that triggers the data interpolation program and is used to start subsequent data fill operations. After receiving the signal, the system will start a pre-trained autoregressive integral moving average model to perform short-term prediction fill on the real-time data within the missing time window.

[0078] Federated Learning Framework: This system employs the classic federated averaging algorithm. Under this framework, the central server first distributes the global model parameters to each production node. Each node independently trains its model using its local data and uploads only the updated model parameters, not the original data, back to the server. The server then performs a weighted average of these parameters to form new global model parameters, which are then distributed again. This process iterates, enabling collaborative training of the global model while protecting data privacy.

[0079] Data imputation model: The autoregressive integral moving average model used can be parameterized based on its historical data to predict short-term future data. The model is expressed as follows: ,in Let the order be the autoregressive order. Let be the difference order. This refers to the order of the moving average. These parameters can be determined through time-series analysis of historical data, such as ACF / PACF plot analysis.

[0080] This measure significantly reduces the global data synchronization latency caused by waiting for delayed data, within an acceptable range of accuracy error.

[0081] Conversely, when The indicator is below or equal to the confidence threshold When the system determines that the current data distribution is unreliable and any form of padding may introduce unpredictable noise data, the module generates a strict wait signal. The strict wait signal is an instruction to suspend data padding and maintain a waiting state. This instruction will cause the system to suspend any data padding operation to ensure the absolute authenticity of the original data until the delayed data packet actually arrives, thereby avoiding low-quality inference data from polluting the global model.

[0082] The addition of this embodiment enables the system to have intelligent and differentiated capabilities in handling data latency issues; it abandons the traditional one-size-fits-all delay waiting or forced filling strategy, and instead makes a dynamic and optimal trade-off between data timeliness and data authenticity based on real-time quantified data quality, thereby maximizing data utilization efficiency while ensuring the stability of the global model.

[0083] The process of generating the model trust index is as follows:

[0084] Obtain the defect prediction recall rate of the global defect prediction model at the current production node;

[0085] By combining the weighting coefficients preset by managers based on the strategic goals of the enterprise at different stages, the defect prediction recall rate and semantic consistency health index are calculated by weighted geometric average to generate a comprehensive model trust index, ensuring that the index is highly sensitive to any decline in predictive ability or data quality.

[0086] This embodiment is a specific implementation of the model trust index generation process in the above-mentioned federal defect prediction and model trust evaluation module; the core is to construct a comprehensive evaluation index that can simultaneously reflect the model prediction performance and its data foundation stability.

[0087] The generation process of the model's trust index is limited; the module obtains the defect prediction recall rate of the global defect prediction model at the current production node. Defect prediction recall rate It refers to the proportion of defect samples correctly predicted by the model to the total number of actual defect samples. It is a well-known standard indicator for evaluating the recall capability of the prediction model. It is derived from the real-time calculation by this module by comparing the model's early warning results with the actual results of production line quality inspection.

[0088] The module combines weighting coefficients preset by managers based on the strategic goals of the company at different stages of development to predict the recall rate of this defect. Semantic consistency health index calculated by the preceding module A weighted geometric mean is calculated to generate a comprehensive model confidence index; the calculation formula is designed as follows:

[0089] in, The model trust index is a score that comprehensively reflects the reliability of the model. It is a floating-point number and is calculated by this formula. Defect prediction recall rate reflects the predictive performance of the model, and the value comes from the real-time monitoring of this module; Semantic consistency health reflects data quality, and its value comes from the calculation results of the data semantic consistency quantification module; , Weighting coefficients are dimensionless preset parameters used to adjust the relative importance of predictive performance and data quality in the final trust assessment. They are set by company managers based on strategic needs; for example, in a phase pursuing ultimate production stability, they can be set to... To increase penalties for data quality issues; in stages requiring rapid verification of new material properties, settings can be configured... To place more emphasis on the model's predictive performance;

[0090] Weighting coefficient , The settings must follow the company's strategic priorities;

[0091] Prioritize data quality: In the initial stages of introducing new supplier materials or during periods when system stability is critical, a system should be designed with data quality in mind. .For example, , The formula then becomes This setting increases the penalty for decreased data quality on the final trust score, prompting the system to respond quickly when there is slight data drift.

[0092] Prioritizing predictive capabilities: In stages requiring rapid verification of new material properties or exploration of new processes, a predictive capability should be set. ;For example, , The formula then becomes This setting focuses more on the actual predictive performance of the model, allowing reliance on the model's predictions even when data quality fluctuates slightly.

[0093] The formula's structure is derived from multi-objective decision-making theory, and the use of geometric mean ensures the reliability index. right or A significant decrease in any one of the indicators demonstrates high sensitivity, reflecting the inherent logic that the overall reliability of the model depends on its weakest link.

[0094] It should be noted that, for simplicity, this formula assumes a multiplicative relationship between data quality and model predictive power; in practical applications, the interaction between the two may be more complex. Future model iterations could consider employing more advanced nonlinear mapping models, such as learning and fitting data through regression analysis or small neural networks. right The true dose-response relationship of the influence is obtained, thereby further improving the physical fidelity of the model confidence index;

[0095] The model trust index generated in this way It transcends traditional model evaluation methods that only focus on prediction accuracy; it establishes a more comprehensive evaluation system, enabling the system to know not only whether the model's predictions are accurate, but also whether the predictions are based on reliable data; this built-in sensitivity to data quality allows the final decision to anticipate and avoid deep-seated risks caused by data pollution earlier.

[0096] Example 3

[0097] The generation process of strategic adjustment factors is as follows:

[0098] Obtain the current prediction accuracy of the plant and the preset failure boundary set by the production process experts, and calculate the difference between the two to determine the safe distance representing the system's buffer space;

[0099] The model's trust index is reverse-processed to obtain the distrust index;

[0100] Dividing the degree of distrust by the safety distance generates a strategic adjustment factor, thereby using the buffer space of the system's distance from failure as a risk amplifier, so that the closer the prediction accuracy is to the failure boundary, the more significant the increase in risk assessment results.

[0101] The final decision and response of the dynamic equilibrium and strategic decision-making module are as follows:

[0102] When the strategic adjustment factor exceeds the preset adjustment factor threshold, an isolation signal is generated, automatically disconnecting the corresponding production node from the federated learning network and instructing it to degrade to local statistical process control mode, sacrificing the short-term advanced prediction function of the node to protect the purity of the global model.

[0103] When the strategic adjustment factor does not exceed the preset adjustment factor threshold, a normal operation signal is generated to maintain the normal training and prediction status of the production node in the federated learning network.

[0104] This embodiment is an implementation of the above-mentioned dynamic balance and strategic decision-making module, and describes in detail the complete process of risk quantification, decision generation and final response.

[0105] As defined above, the generation process of strategic adjustment factors in this module is specified; the module obtains the real-time forecast accuracy of the current plant. Prediction accuracy This refers to the proportion of samples correctly predicted by the model out of the total number of samples, which is obtained from the real-time monitoring of this module; at the same time, the module obtains the preset failure boundary set by production process experts according to the quality control red line. Preset failure boundary This refers to the minimum acceptable level of system performance, such as 70% accuracy. Its purpose is to define an inviolable red line, derived from the regulations of the production process and quality management departments. The module calculates the difference between these two values. To determine the safe distance representing the current buffer space of the system;

[0106] The module will use the model trust index calculated in the previous steps. Perform reverse processing to calculate the level of distrust. ;

[0107] The module divides the level of distrust by the safe distance to generate a strategy adjustment factor. This process uses the system's buffer space from failure as a risk amplifier; the calculation formula is designed as follows:

[0108] in, The strategic adjustment factor is an indicator for quantifying systemic risk. The higher the value, the higher the risk. It is calculated by this formula and is a dimensionless indicator used to quantify risk, rather than a quantity with a specific physical unit. Model trust score, the value of which comes from the Federal Flaw Prediction and Model Trust Assessment module; : Current prediction accuracy, the value is derived from real-time monitoring of this module; Preset failure boundary is a parameter preset in the system; as well as It is a percentage or proportion, and is dimensionless; the mathematical structure of the formula makes it possible to predict accuracy. Approaching the failure boundary When the denominator approaches zero, it leads to The value is amplified dramatically, resulting in a non-linear increase in the risk assessment results, which is extremely sensitive to states on the verge of failure. It is a very small positive number, which is a dimensionless correction term, for example, 0.01, used to prevent the denominator from being zero and to smooth out the fraction when... near The risk amplification effect at that time;

[0109] To more comprehensively quantify systemic risk, this system introduces an additional risk correction term in the calculation of the strategic adjustment factor. The revised formula is as follows:

[0110] in, The external risk correction factor is a dimensionless multiplicative factor, which is a multiplier greater than or equal to 1. Comprehensive consideration was given to factors such as network stability metrics, for example, when the packet loss rate exceeded a threshold in the past hour. The health status of key sensors, for example, when a sensor malfunction is detected by a self-diagnostic algorithm, The specific values ​​and calculation methods of these correction factors, including multiple external variables, can be determined by conducting retrospective analysis of historical abnormal events and establishing corresponding mapping rules to ensure that they can more accurately reflect the complex risks in reality.

[0111] As defined above, the final decision-making and response mechanism of this module is specified; the module will calculate the strategic adjustment factors in real time. With preset adjustment factor threshold Perform comparison; preset adjustment factor threshold This refers to the critical risk value used to trigger system state switching. Its function is to serve as a decision criterion for activating or deactivating the isolation mechanism. It is derived from offline simulations using massive amounts of historical data, testing the long-term overall production efficiency of the system under different threshold settings, and ultimately selecting the value that maximizes long-term benefits. value;

[0112] Preset failure boundary The setting needs to be based on rigorous quantitative analysis of historical data; for example, by retrospectively analyzing historical cases of production stoppages due to production accidents or serious quality problems, the lowest historical value of prediction accuracy before these events occurred can be determined. Using this lowest value as a benchmark, and applying a safety margin on top of it, the prediction accuracy can be set. For example, if the historical lowest accuracy rate is 75%, then... Set to 80%;

[0113] Adjusting factor threshold Determining the optimal solution is a system-level optimization problem, which cannot be calculated using a single formula. Instead, it requires large-scale offline simulation experiments to find the optimal solution. The method is as follows:

[0114] Data playback: Simulations are performed using historical production data from several months or even several years to simulate various abnormal situations such as material changes and data delays;

[0115] Threshold test: for Take a series of candidate values, such as And in each simulation, when When the threshold is exceeded, the isolation mechanism is triggered;

[0116] Benefit assessment: Comparison of different The long-term production benefits during the simulation process, such as the net benefit between production losses caused by mis-isolation and global model failure losses caused by non-isolation;

[0117] Optimal choice: The final choice that maximizes the system's long-term net benefit. value;

[0118] When the strategic adjustment factor of a certain production node Exceeding the preset adjustment factor threshold When the module determines that the node is too risky and poses a threat to the global model, it generates an isolation signal. The isolation signal is a system instruction that triggers the node to disconnect and downgrade. This instruction will automatically disconnect the corresponding production node from the federated learning network and instruct it to downgrade to a statistical process control mode that relies solely on local historical data. This action sacrifices the node's short-term advanced prediction capabilities to protect the purity and stability of the global model.

[0119] When strategic adjustment factors The preset adjustment factor threshold has not been exceeded. At this time, the module generates a normal operation signal; the normal operation signal refers to the instruction to maintain the current network connection and operation mode; this signal will maintain the normal training and prediction state of the production node in the federated learning network;

[0120] To quantify net benefits, this system defines a comprehensive benefit function. This function is calculated through backtracking analysis, specifically as follows:

[0121] in, : The benefit of avoiding scrap loss due to accurate model prediction, which can be calculated by multiplying the number of defects that the model successfully warned of and avoided in normal operation mode by the value per unit of scrap in historical data.

[0122] Production losses caused by mis-isolation, for example, in Exceed However, when the actual data quality and model performance are within acceptable limits, the production interruption loss caused by temporarily disconnecting a node can be calculated by multiplying the isolation duration by the production value per unit time.

[0123] Loss due to global model failure caused by failure to isolate high-risk nodes, for example, noisy data introduced by the node pollutes the global model, causing a decrease in the prediction accuracy of other nodes. This cost can be calculated by assessing the additional scrap rate increase caused by the deterioration of global model performance.

[0124] The choice makes Maximize The value serves as the optimal threshold for the system;

[0125] This embodiment introduces a strategic adjustment factor. Its decision-making mechanism endows the system with the highest level of autonomous risk avoidance capability; instead of passively monitoring data quality or model performance, it proactively and forward-lookingly quantifies the failure risk of the entire node; by decisively executing isolation operations when the risk is drastically amplified, the system can protect itself before systemic failure occurs, sacrificing the short-term functional degradation of a single node in exchange for the long-term stable, efficient and healthy operation of the entire cross-regional production network, and achieving the optimization of global interests;

[0126] Before system deployment, strategic adjustment factors need to be considered. The calculation formula underwent a comprehensive robustness check, which included simulations and tests for the following extreme cases:

[0127] Prediction accuracy Approaching the preset failure boundary Verify whether the system can quickly and stably trigger the isolation signal when the denominator is close to zero, and check whether there is a system crash due to numerical overflow.

[0128] Model confidence Approaching 0 or 1: Validation when data quality or predictive performance is extremely poor or extremely good. Whether the calculations and system responses are logically consistent, for example, when When it approaches 0, The value should approach infinity, and the system should unconditionally trigger isolation.

[0129] Boundary tests within the normal value range: and Perform cross-combination tests within the upper and lower limits of its normal fluctuation range to ensure that all possible inputs are within acceptable limits. The values ​​are all within a controllable range and can be compared with the preset threshold. Produce reasonable comparison results;

[0130] These tests will ensure that the mathematical behavior of the core decision model is consistent with the expected physical behavior of the system, thereby enhancing the reliability of the entire system.

[0131] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An in-swinging casement assembly production line full-process monitoring data management system, characterized in that, The application relates to a system and method for dynamically balancing the quality of data and the accuracy of a model. The system comprises: a distributed data acquisition and preprocessing module for acquiring sensor data on a production line at each production node and attaching high-precision timestamps and material supplier identifiers; a data semantic consistency quantification module for calculating the KL divergence value of the current sensor data distribution relative to a preset reference data distribution model, and generating a quantified semantic consistency health index by using an exponential function for mapping processing, so as to accurately evaluate the data quality change caused by the mixing use of new and old supplier materials; a federal defect prediction and model trust evaluation module for combining the semantic consistency health index and the defect prediction recall rate of a global defect prediction model, and generating a comprehensive model trust index by using a weighted geometric mean calculation, so as to synchronously reflect the double guarantee of the prediction ability and the data quality; a dynamic balance and strategic decision module for calculating the safety distance between the current prediction accuracy and a preset failure boundary, combining the model trust index to calculate the distrust index, and then generating a strategic adjustment factor to quantify the system risk; and comparing the strategic adjustment factor with a preset threshold to generate an isolation signal for triggering an isolation mechanism or a normal operation signal for maintaining the current state. The generation process of the strategic adjustment factor is as follows: obtaining the prediction accuracy of the current factory and the preset failure boundary set by the production process expert, calculating the difference between the two to determine the safety distance representing the system buffer space; performing reverse processing on the model trust index to obtain the distrust index; 2. The full-process monitoring data management system for an interior opening window assembly production line according to claim 1, wherein, dividing the distrust index by the safety distance to generate the strategic adjustment factor. The generation process of the semantic consistency health index is as follows: obtaining a reference data distribution model constructed based on high-quality single supplier material historical data, and obtaining a current experience probability distribution constructed based on current sensor data; calculating the KL divergence value of the current experience probability distribution relative to the reference data distribution model, so as to quantify the deviation degree of the current data distribution; 3. The full-process monitoring data management system for an interior opening window assembly production line according to claim 1, wherein, using an exponential function containing a preset attenuation coefficient to perform mapping processing on the KL divergence value, generating the semantic consistency health index, and thus converting the unbounded divergence value into an intuitive and standardized data quality score. The adaptive data synchronization and time delay compensation module has a decision processing process as follows: comparing the semantic consistency health index with a confidence threshold determined through offline simulation experiments; when the semantic consistency health index is higher than the confidence threshold, determining that the data source quality is stable and generating a prediction filling signal to start a pre-trained autoregressive integral moving average model for data filling; 4. The full-process monitoring data management system for an interior opening window assembly production line according to claim 1, wherein, when the semantic consistency health index is lower than or equal to the confidence threshold, determining that the data distribution is unreliable and generating a strict waiting signal to suspend any data filling operation. The generation process of the model trust index is as follows: obtaining the defect prediction recall rate of a global defect prediction model at the current production node; 5. The full-process monitoring data management system for an interior opening window assembly production line according to claim 1, wherein, combining the weight coefficient preset by the manager according to the strategic target of the enterprise at different stages, and performing weighted geometric mean calculation on the defect prediction recall rate and the semantic consistency health index, so as to generate a comprehensive model trust index. The final decision and response of the dynamic balance and strategic decision module are as follows: When the strategic adjustment factor exceeds the preset adjustment factor threshold, an isolation signal is generated, the corresponding production node is automatically disconnected from the federated learning network, and a downgrade to a local statistical process control mode is instructed to run; When the strategic adjustment factor does not exceed the preset adjustment factor threshold, a normal operation signal is generated, and the normal training and prediction state of the production node in the federated learning network is maintained.

Images