An intelligent quality discrimination and decision method based on an industrial internet

Abstract

The application discloses an intelligent quality discrimination and decision method based on an industrial internet, relates to the technical field of intelligent manufacturing and quality predictive maintenance, and comprises the following steps: acquiring real-time edge credibility scores corresponding to key index pairs constituting a latest dynamic correlation boundary; calculating effective deviation degrees of the index value vectors to the boundary based on the acquired real-time edge credibility scores; and generating a comprehensive risk value R based on the effective deviation degrees and a fault mode severity weight W associated with the dynamic correlation boundary. The application realizes double-layer detection of single-index overrun and multi-index correlation imbalance by constructing a composite decision model containing a static threshold and a dynamic correlation boundary, and enhances the identification capability for complex fault modes and the accuracy of early warning.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing and predictive maintenance technology, and in particular to an intelligent quality anomaly detection and decision-making method based on the Industrial Internet. Background Technology

[0002] With the deep integration of industrial internet, big data, and artificial intelligence technologies, the manufacturing industry is rapidly evolving towards intelligence, adaptability, and high reliability. Real-time collection of multi-dimensional data from the entire production process, and the rapid and accurate identification of quality anomalies to drive autonomous decision-making and optimization, has become a key to achieving intelligent manufacturing, ensuring product consistency, and improving production efficiency. This trend presents new challenges to traditional quality monitoring models, requiring them to transform from a lagging, isolated, and experience-dependent approach into a real-time, interconnected, and self-learning intelligent system.

[0003] Currently, quality monitoring and anomaly detection in industrial settings primarily rely on two technical approaches: one is a real-time alarm system based on preset thresholds, which sets fixed safety boundaries for individual process parameters. An alarm is triggered when the monitored value exceeds these boundaries. This method responds quickly but cannot reveal the complex coupling relationships between multiple parameters or early-stage hidden faults. The other approach is post-hoc analysis based on statistical process control, which uses periodic sampling and tools such as control charts for offline diagnosis. While this method can identify certain statistical patterns, it suffers from significant lag and cannot meet the needs of real-time intervention. Furthermore, existing advanced data analysis solutions often directly process raw collected data, lacking systematic evaluation and cleaning mechanisms for the data's reliability. When faced with common industrial problems such as communication interference, signal fluctuations, and data loss, the accuracy of the analysis and the reliability of the decisions cannot be guaranteed.

[0004] Existing technologies suffer from the following prominent problems: First, the methods for identifying anomalies are limited and fail to effectively integrate the advantages of rapid response from single-point thresholds with in-depth insights from multi-indicator correlations, resulting in insufficient early identification capabilities for complex fault modes. Second, the system lacks adaptability; thresholds and rules are typically set statically, making it difficult to dynamically adjust them in response to equipment status, process evolution, and environmental changes, which can easily lead to false alarms or missed alarms. Third, the foundation for data-driven decision-making is weak; a quality and reliability assessment and transmission mechanism that runs through the entire chain from data acquisition and processing to decision-making has not been established, leading to increased decision-making risks when data quality fluctuates.

[0005] Therefore, it is essential to invent an intelligent quality discrimination and decision-making method based on the Industrial Internet to solve the above problems. Summary of the Invention

[0006] The purpose of this invention is to provide an intelligent quality discrimination and decision-making method based on the Industrial Internet to solve the problems mentioned in the background art.

[0007] To achieve the above objectives, the present invention provides the following technical solution: an intelligent quality anomaly detection and decision-making method based on the Industrial Internet, specifically including the following steps:

[0008] S1. Through an intelligent gateway with adaptive protocol parsing, collect real-time data of at least three interrelated quality indicators in the production unit, and generate an initial credibility score Ci(t) for each data point based on multi-dimensional verification during the collection process. This score is attached to the corresponding data point as the initial value of the credibility label.

[0009] S2. Based on the initial confidence score Ci(t) carried by the confidence label, perform real-time patching and lightweight statistical process control screening on the data stream at the edge side based on time series prediction; for the patched data points, update their confidence label values ​​to the edge confidence score Ce(t); for the data that passes the screening, append an "edge screening passed" flag; finally, output the preprocessed data stream with the updated confidence label and screening flag;

[0010] S3. Based on the historical preprocessed data stream, construct a composite decision model, which includes a static threshold boundary and at least one dynamic correlation boundary, wherein each dynamic correlation boundary is pre-assigned a boundary type identifier representing a specific fault mode and a fault mode severity weight W.

[0011] S4. For the preprocessed data stream uploaded in real time, extract its indicator value vector and execute:

[0012] First-level judgment: Determine whether the indicator value vector exceeds the static threshold boundary;

[0013] Second-level judgment: If not exceeded, proceed with the following steps:

[0014] S41. Obtain the real-time edge credibility score corresponding to the key indicators that constitute the nearest dynamic association boundary.

[0015] S42. Based on the obtained real-time edge credibility score, calculate the effective deviation of the index value vector from the boundary;

[0016] S43. Based on the effective deviation and the fault mode severity weight W associated with the dynamic correlation boundary, generate a comprehensive risk value R;

[0017] S5. Predefine multiple control action levels and set a comprehensive risk threshold for each level; based on the comprehensive risk value R output by the real-time decision-making stage and the boundary type identifier of the dynamic associated boundary it triggers, select the control action corresponding to the highest level that its comprehensive risk value R exceeds and execute it.

[0018] S6. Collect the effect data after the control action is executed, and optimize the confidence interval width of the dynamic association boundary, the fault mode severity weight W, and the comprehensive risk threshold of the control action level through reinforcement learning.

[0019] Preferably, generating the initial confidence score Ci(t) in S1 includes:

[0020] The device communication protocol is automatically identified by the protocol probe, and the corresponding parser is loaded.

[0021] The system performs instantaneous value jump detection, message continuity verification, and logical consistency verification between numerical values ​​and device status in parallel to obtain a comprehensive verification result.

[0022] Based on the comprehensive verification results, the initial confidence score Ci(t) is calculated.

[0023] Preferably, the update process of the credibility tag in S2 is as follows:

[0024] For data points in the credibility label whose initial credibility score Ci(t) is lower than the first threshold, a repair value is generated based on their historical sequence using a prediction model.

[0025] The credibility label value of the successfully repaired data points is updated to Ce(t)=k*Ci(t)+(1-k)*Cp, where Cp is the preset base credibility of the repair operation, Ci(t) is the initial credibility score, and k is the weight coefficient.

[0026] Preferably, the process of constructing the dynamic association boundary includes:

[0027] Preprocessed data stream segments that consistently bear the "edge screening passed" label and are marked as normal operating conditions are selected from historical data to form a highly reliable training sample set;

[0028] The implicit correlation function f(xi, xj, ...) = 0 was obtained by fitting the data using a multivariate nonlinear regression method with L2 regularization.

[0029] Calculate the residual distribution from the training samples to the function f=0, and define the confidence interval width using the predetermined higher quantile of this distribution to form a dynamic correlation boundary.

[0030] Preferably, the second-level judgment includes:

[0031] Calculate the shortest geometric distance D from the index value vector X to the nearest dynamic association boundary;

[0032] Identify the key indicator pairs (xm, xn) that constitute the boundary and obtain their corresponding real-time edge confidence scores Cem and Cen;

[0033] The effective deviation is calculated using the formula Deff=D×(1+γ / (min(Cem,Cen)+ε)), where γ is the sensitivity coefficient and ε is the decimal point to prevent zeroing.

[0034] The overall risk value R is calculated using the formula R=W×(Deff / Dref), where Dref is the reference distance constant and W is the severity weight of the failure mode.

[0035] Preferably, the control action levels include at least early warning, process parameter fine-tuning, production step suspension, and overall line shutdown; when executing control actions, the boundary type identifier of the triggered dynamic associated boundary is output synchronously as fault pre-diagnosis information.

[0036] Preferably, the optimization in S6 includes: after each control action is executed, starting an effect observation window to collect subsequent quality indicator data and their credibility labels;

[0037] Analyze the data within the observation window. If the indicator value vector regresses to the normal decision region and remains stable, it is marked as an effective intervention; otherwise, it is an intervention to be optimized.

[0038] Data containing event sequences and intervention effect labels are used as training samples to adjust the confidence interval width of the dynamic association boundary that triggers the intervention, the failure mode severity weight W, and the comprehensive risk threshold for increasing or decreasing the level of its associated control actions.

[0039] The technical effects and advantages of this invention are as follows:

[0040] 1. This invention introduces a full-process credibility scoring and labeling mechanism from data acquisition to edge processing, thereby achieving dynamic evaluation and improvement of industrial data quality, effectively reducing the risk of misjudgment caused by data anomalies or noise, and improving the reliability of anomaly judgment decisions.

[0041] 2. This invention constructs a composite decision model that includes static thresholds and dynamic correlation boundaries, thereby achieving dual-layer detection of single indicator exceeding limits and multi-indicator correlation imbalance, enhancing the ability to identify complex fault modes and the accuracy of early warning.

[0042] 3. This invention calculates the effective deviation by combining real-time reliability scoring and integrates the severity weight of fault modes to generate a comprehensive risk value, thereby realizing intelligent risk quantification based on data reliability and fault severity, and supporting graded and precise control action triggering.

[0043] 4. This invention collects effect data after executing control actions and continuously optimizes boundary parameters, weights, and risk thresholds through reinforcement learning, forming a closed-loop self-learning mechanism of "decision-execution-feedback-optimization", which improves the system's adaptability and the level of intelligence in long-term operation and maintenance. Attached Figure Description

[0044] Figure 1 This is a flowchart of the overall method of the present invention.

[0045] Figure 2 This is a flowchart illustrating the credibility score generation and updating process of the present invention.

[0046] Figure 3 This is a flowchart of the secondary judgment and risk calculation process of the present invention.

[0047] Figure 4 This is a flowchart illustrating the control action execution of the present invention.

[0048] Figure 5 The following is a flowchart of the feedback optimization closed-loop process of this invention. Detailed Implementation

[0049] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0050] This invention provides, for example Figure 1 The intelligent quality anomaly detection and decision-making method based on the Industrial Internet, as shown, specifically includes the following steps:

[0051] S1. Through an intelligent gateway with adaptive protocol parsing, collect real-time data of at least three interrelated quality indicators in the production unit, and generate an initial credibility score Ci(t) for each data point based on multi-dimensional verification during the collection process. This score is attached to the corresponding data point as the initial value of the credibility label.

[0052] Furthermore, in the above technical solution, refer to Figure 2 The generation of the initial credibility score Ci(t) in S1 includes:

[0053] The device communication protocol is automatically identified by the protocol probe, and the corresponding parser is loaded.

[0054] The system performs instantaneous value jump detection, message continuity verification, and logical consistency verification between numerical values ​​and device status in parallel to obtain a comprehensive verification result.

[0055] Based on the comprehensive verification results, the initial confidence score Ci(t) is calculated.

[0056] It is important to understand that the calculation of the initial credibility score Ci(t) is the primary step in data quality assurance. The protocol probe automatically identifies the device communication protocol by monitoring network packet header features and matching them with a pre-defined protocol fingerprint database. It then dynamically loads the corresponding parser to decode the packet content, extracting the quality indicator values, timestamps, and source device identifiers. The instantaneous value jump detection calculates the absolute difference between the current value and the previous valid value and compares it to a dynamic threshold calculated based on historical data of that indicator. If the difference exceeds the dynamic threshold, it is considered a suspicious jump. The packet continuity verification checks whether the sequence number or timestamp of the data packets is continuous and calculates the packet loss rate within a specific time window. The numerical value and device state logic consistency verification determines whether the currently reported value is within a reasonable range allowed by the device's current operating state (e.g., standby, processing, maintenance) according to predefined mapping rules between the device state machine and the numerical range. The comprehensive verification result consists of sub-results from the above three verification dimensions. One specific implementation of calculating the initial credibility score Ci(t) based on the comprehensive verification results is as follows: base scores are assigned to instantaneous value jump detection, message continuity verification, and logical consistency verification respectively. When a certain verification passes, all scores are obtained; when some are suspicious, points are deducted according to preset rules; and when completely abnormal, zero scores are obtained. The three scores are weighted and summed, and then mapped to the interval between 0 and 1 to obtain the initial credibility score Ci(t).

[0057] To illustrate the process of generating the initial credibility score in more detail, the following table shows:

[0058]

[0059] S2. Based on the initial confidence score Ci(t) carried by the confidence label, perform real-time patching and lightweight statistical process control screening on the data stream at the edge side based on time series prediction; for the patched data points, update their confidence label values ​​to the edge confidence score Ce(t); for the data that passes the screening, append an "edge screening passed" flag; finally, output the preprocessed data stream with the updated confidence label and screening flag;

[0060] Furthermore, in the above technical solution, refer to Figure 2 The update process of the credibility label in S2 is as follows:

[0061] For data points in the credibility label whose initial credibility score Ci(t) is lower than the first threshold, a repair value is generated based on their historical sequence using a prediction model.

[0062] The credibility label value of the successfully repaired data points is updated to Ce(t)=k*Ci(t)+(1-k)*Cp, where Cp is the preset base credibility of the repair operation, Ci(t) is the initial credibility score, and k is the weight coefficient.

[0063] It is important to understand that the core of the real-time patching based on time series prediction lies in the reasonable estimation and replacement of low-confidence data points. One specific implementation of the prediction model is to use an autoregressive integral moving average model or a long short-term memory network model. During patching, the historical numerical sequence of the quality index to which the data point belongs, stored on the edge side, within the most recent time window, is extracted. After removing outliers marked as low-confidence, the cleaned sequence is input into the prediction model, which outputs a predicted value for the current moment, which is then used as the patched value for that data point.

[0064] The lightweight statistical process control screening refers to online calculation of the real-time data stream of each quality indicator at the edge to determine whether it is in a statistically controlled state. One specific implementation involves maintaining a dynamic moving average and moving standard deviation for each indicator, and calculating real-time upper and lower control limits based on the principle of "mean ± 3 times standard deviation." For newly arriving data points, if their values ​​fall between the upper and lower control limits, they are considered to have passed the screening, and a "edge screening passed" flag is added. Simultaneously, the moving average and moving standard deviation are updated using this new data point. The screening and patching can be performed in parallel without affecting each other.

[0065] Regarding the update of the credibility label, the first threshold is a pre-set empirical value, such as 0.6, used to quickly filter out data points with low credibility that need to be repaired first. The weighting coefficient k is used to balance the influence of the original data credibility and the credibility of the repair operation on the final score. Its value ranges from 0 to 1 and can be set to a fixed value (such as 0.3) based on business experience, or dynamically adjusted according to the confidence level of the prediction model's prediction result. The basic credibility Cp of the repair operation represents the basic trust in the system's repair capability and is a preset constant, usually set to a value between 0.5 and 0.8, such as 0.7. The meaning of the update formula Ce(t)=k*Ci(t)+(1-k)*Cp is that the final credibility of the repaired data point is obtained by weighted summation of the residual credibility of the original data and the credibility introduced by the current repair operation.

[0066] To visually illustrate the edge confidence score update and screening process, the following table is provided:

[0067]

[0068] S3. Based on the historical preprocessed data stream, construct a composite decision model, which includes a static threshold boundary and at least one dynamic correlation boundary, wherein each dynamic correlation boundary is pre-assigned a boundary type identifier representing a specific fault mode and a fault mode severity weight W.

[0069] Furthermore, in the above technical solution, the process of constructing the dynamic association boundary includes:

[0070] Preprocessed data stream segments that consistently bear the "edge screening passed" label and are marked as normal operating conditions are selected from historical data to form a highly reliable training sample set;

[0071] The implicit correlation function f(xi, xj, ...) = 0 was obtained by fitting the data using a multivariate nonlinear regression method with L2 regularization.

[0072] Calculate the residual distribution from the training samples to the function f=0, and define the confidence interval width using the predetermined higher quantile of this distribution to form a dynamic correlation boundary.

[0073] It is important to understand that the construction of the dynamic correlation boundary is an offline modeling process based on historical data, aimed at uncovering the inherent normal correlations between multiple indicators. "Continuously bearing the 'edge screening passed' marker" means that within a preset time segment length (e.g., 5 consecutive minutes), the "edge screening passed" marker exists for all data points in that segment. "Marked as normal operating condition" means that the production time periods corresponding to these historical data segments are explicitly recorded in the production logs or quality records of the production unit as a normal production stage with no faults and qualified output.

[0074] The highly reliable training sample set consists of a large number of data segments that meet the above conditions and contain at least three interrelated quality indicator values. Each sample point is a multi-dimensional vector, and its dimensions correspond to the interrelated quality indicators.

[0075] The purpose of "fitting using a multivariate nonlinear regression method with L2 regularization" is to find an implicit mathematical model that can describe the stable relationship between these quality indicators under normal operating conditions. One specific implementation is as follows: Assume the correlation function f is a polynomial function of the input indicators. By substituting the indicator values ​​from the training samples into a predefined polynomial structure, a system of equations with polynomial coefficients as unknowns is constructed. The goal of the fitting is to find a set of coefficients such that for all highly reliable training samples, the output value of the function f(xi, xj, ...) is as close to zero as possible. This is achieved by minimizing a loss function, defined as the sum of f(xi, xj, ...)^2 over all training samples plus an L2 regularization term (i.e., the sum of squares of all coefficients multiplied by a regularization coefficient λ). The role of L2 regularization is to prevent overfitting and improve its generalization ability. Solving this optimization problem yields the specific coefficients of the implicit correlation function f(xi, xj, ...) = 0.

[0076] After obtaining the implicit function f, the residuals from the training samples to the function f=0 are calculated. For each training sample (index value vector), it is substituted into the function f, and the absolute value of the calculated function value f is regarded as the "residual" or "deviation" from the implicit association of that sample. The residuals of all highly reliable training samples are collected to form a residual distribution.

[0077] Defining the confidence interval width using a predetermined higher quantile of the distribution means selecting a high quantile of the residual distribution, such as the 95th or 99th quantile, and denoting the residual value corresponding to that quantile as δ. The "dynamic correlation boundary" is thus defined as the spatial region formed by all index value vectors satisfying the inequality |f(xi, xj, ...)| ≤ δ. This boundary is not a simple threshold, but a complex boundary characterizing the normal correlation between multiple indices, jointly defined by implicit functions and confidence intervals. Each such dynamic correlation boundary is associated with a potential, specific multi-indicator imbalance mode and is therefore assigned a unique boundary type identifier and a failure mode severity weight W.

[0078] S4. For the preprocessed data stream uploaded in real time, extract its indicator value vector and execute:

[0079] First-level judgment: Determine whether the indicator value vector exceeds the static threshold boundary;

[0080] Second-level judgment: If not exceeded, proceed with the following steps:

[0081] S41. Obtain the real-time edge credibility score corresponding to the key indicators that constitute the nearest dynamic association boundary.

[0082] S42. Based on the obtained real-time edge credibility score, calculate the effective deviation of the index value vector from the boundary;

[0083] S43. Based on the effective deviation and the fault mode severity weight W associated with the dynamic correlation boundary, generate a comprehensive risk value R;

[0084] Furthermore, in the above technical solution, refer to Figure 3 The second level of judgment includes:

[0085] Calculate the shortest geometric distance D from the index value vector X to the nearest dynamic association boundary;

[0086] Identify the key indicator pairs (xm, xn) that constitute the boundary and obtain their corresponding real-time edge confidence scores Cem and Cen;

[0087] The effective deviation is calculated using the formula Deff=D×(1+γ / (min(Cem,Cen)+ε)), where γ is the sensitivity coefficient and ε is the decimal point to prevent zeroing.

[0088] The overall risk value R is calculated using the formula R=W×(Deff / Dref), where Dref is the reference distance constant and W is the severity weight of the failure mode.

[0089] It's important to understand that step S4 is a real-time decision-making phase executed in the cloud or on a central server. The "static threshold boundary" is a pre-set, absolute safety limit that cannot be exceeded for each individual quality indicator, typically based on equipment physical limits, process safety specifications, or stringent product quality standards. The first-level judgment checks whether any indicator value in the real-time indicator value vector exceeds its corresponding static threshold boundary. If it does, it is directly judged as a serious anomaly, immediately triggering the highest level of control action (such as shutting down the entire production line) without proceeding to the subsequent second-level judgment.

[0090] If the first-level judgment indicates that no extreme anomalies have occurred in the real-time data, then the second-level judgment is initiated, which involves a more refined risk assessment based on the dynamic correlation boundaries in the composite decision model. The "nearest dynamic correlation boundary" refers to the dynamic correlation boundary in the decision space composed of multiple dynamic correlation boundaries that is geometrically closest to the real-time indicator value vector. A specific way to calculate this closest distance is as follows: for each dynamic correlation boundary (i.e., the region defined by |f(xi, xj, ...)|≤δ), calculate the absolute value of the function value f(X) obtained by substituting the real-time indicator value vector X into its implicit function f. This absolute value can be considered as a measure of X's distance to the boundary center (f=0); the dynamic correlation boundary that minimizes this absolute value is the "nearest" dynamic correlation boundary.

[0091] The step S41, "obtaining the real-time edge credibility score corresponding to the key indicator pair constituting the most recent dynamic association boundary," specifically involves the following: When each dynamic association boundary is constructed, its implicit function f(xi, xj, ...) = 0 reveals which indicators are most sensitive to changes in function values. By analyzing the partial derivatives of function f with respect to each indicator, the two indicators that contribute the most to |f(X)| at the current real-time indicator value X can be identified, denoted as (xm, xn), which are the "key indicator pair." Subsequently, from the preprocessed data stream uploaded in real time, the data points corresponding to these two key indicators xm and xn at the current moment are searched, and their real-time edge credibility scores are read, denoted as Cem and Cen, respectively.

[0092] The core of risk quantification is "calculating the effective deviation of the indicator value vector from the boundary" in S42. First, the shortest geometric distance D from the indicator value vector X to the nearest dynamically related boundary is calculated. In practice, when the boundary is defined as |f(X)|≤δ, if |f(X)|>δ, the distance D can be approximated as |f(X)|-δ; if |f(X)|≤δ, the distance D is zero (indicating it is within the boundary). Then, combining the credibility scores of the acquired key indicator pairs, the effective deviation Deff is calculated using the formula Deff=D×(1+γ / (min(Cem,Cen)+ε)). The key advantage of this formula is that when the credibility of the key indicator data is low (i.e., min(Cem,Cen) is small), the value of the second part of the formula (γ / (min(Cem,Cen)+ε)) increases, thus amplifying the effective deviation Deff. This reflects the decision-making logic that "anomaly judgments based on low-confidence data are inherently riskier and require greater caution," where γ is a sensitivity coefficient used to adjust the strength of the influence of confidence, and ε is a very small positive number (such as 1e-5) to prevent division by zero errors.

[0093] In step S43, "generating a comprehensive risk value R based on the effective deviation and the severity weight W of the fault mode associated with the dynamic correlation boundary" specifically involves calculating R using the formula R = W × (Deff / Dref). Here, W is the pre-set severity weight of the fault mode for the dynamic correlation boundary; a higher weight indicates a more severe consequence of the fault mode corresponding to that boundary. Deff is the effective deviation calculated in the previous step. Dref is a reference distance constant used to normalize the effective deviation; its value can be set as the standard deviation of the residual distribution used when constructing the dynamic correlation boundary, or an empirically set fixed value. The comprehensive risk value R is a dimensionless numerical value that comprehensively reflects the degree to which the current indicator vector deviates from the normal correlation pattern (Deff), the reliability of the data on which this deviation is based (implied in the Deff calculation), and the severity (W) of the fault that this deviation may point to. A larger R value indicates a higher risk currently faced by the system, and the level of control action required may also be higher.

[0094] The following table illustrates the process of secondary judgment and risk calculation:

[0095]

[0096] S5. Predefine multiple control action levels and set a comprehensive risk threshold for each level; based on the comprehensive risk value R output by the real-time decision-making stage and the boundary type identifier of the dynamic associated boundary it triggers, select the control action corresponding to the highest level that its comprehensive risk value R exceeds and execute it.

[0097] Furthermore, in the above technical solution, refer to Figure 4 The control action levels include at least early warning, process parameter fine-tuning, production step suspension, and overall line shutdown; when executing control actions, the boundary type identifier of the triggered dynamic associated boundary is output synchronously as fault pre-diagnosis information.

[0098] It is important to know that the comprehensive risk thresholds form an increasing sequence: T warning <T adjust <T pause <T stop For example, T can be set. warning =1.0, T adjust =2.5, T pause =4.0, T stop =6.0. These thresholds can be initialized based on historical experience or the debugging results during the system trial operation phase.

[0099] Based on the comprehensive risk value R output in the real-time decision-making stage and the boundary type identifier of the dynamically associated boundary it triggers, the control action corresponding to the highest level exceeded by the comprehensive risk value R is selected and executed. The specific logic is as follows: The comprehensive risk value R calculated in real time is compared with the above threshold sequence. If R>=T stop If T pause <=R <T stop If T..., then execute the "Production Step Pause" action; if T... adjust <=R <T pause If T warning <=R <T adjust If R <T warning If the risk is deemed acceptable, no active control actions will be taken, and only routine data recording will be performed.

[0100] The specific implementation methods of the control action include:

[0101] Warning: Send alarm messages to the monitoring terminals or mobile devices of production management personnel, and highlight the triggered boundary type identifier and the current risk value R in the dashboard system.

[0102] Process parameter fine-tuning: Based on the fault mode knowledge base associated with the triggered dynamic association boundary, a preset process parameter adjustment strategy is retrieved (e.g., the set point of a certain temperature control zone is lowered), and adjustment instructions are sent to the corresponding programmable logic controller or industrial controller through the control network.

[0103] Production step pause: Send a command to the manufacturing execution system or production line main controller to pause the equipment or unit currently performing a specific production step, pending manual inspection.

[0104] Complete line shutdown: Sends an emergency shutdown signal to the production line safety control system, triggering a complete safety shutdown process.

[0105] When executing control actions, the boundary type identifier of the triggered dynamic associated boundary is output synchronously as fault pre-diagnosis information. The "boundary type identifier" is a unique code that corresponds one-to-one with the specific dynamic associated boundary constructed in step S3, and the fault mode description, possible causes, and preliminary handling suggestions represented by this identifier are pre-stored in the knowledge base. This identifier, along with the control action command, timestamp, risk value R, key indicator values, and their reliability score, is packaged into a complete decision log and output to the database, monitoring interface, and maintenance work order system, providing direct diagnostic clues for subsequent manual intervention and cause tracing.

[0106] S6. Collect the effect data after the control action is executed, and optimize the confidence interval width of the dynamic association boundary, the fault mode severity weight W, and the comprehensive risk threshold of the control action level through reinforcement learning.

[0107] Furthermore, in the above technical solution, refer to Figure 5 The optimization in S6 includes: after each control action is executed, starting an effect observation window to collect subsequent quality indicator data and its credibility label;

[0108] Analyze the data within the observation window. If the indicator value vector regresses to the normal decision region and remains stable, it is marked as an effective intervention; otherwise, it is an intervention to be optimized.

[0109] Data containing event sequences and intervention effect labels are used as training samples to adjust the confidence interval width of the dynamic association boundary that triggers the intervention, the failure mode severity weight W, and the comprehensive risk threshold for increasing or decreasing the level of its associated control actions.

[0110] It is important to understand that step S6 is a crucial closed-loop link in achieving system self-learning and continuous optimization. The "effect observation window" is a preset time period immediately following the execution of the control action, its length depending on the response characteristics of the production process and the level of the control action. For example, for "early warning" and "process parameter fine-tuning," the observation window might be set to 10-30 subsequent production cycles or 5-15 minutes; for "production step pause" and "complete line shutdown," the observation window might cover dozens of cycles after the production line resumes operation. Within the observation window, the system continuously collects the quality indicator data and their reliability labels output by the production unit, processed through steps S1 and S2, forming an effect evaluation dataset.

[0111] The core of analyzing data within the observation window is to assess whether control actions have brought the production process back on track. Specifically:

[0112] "Indicator value vector regresses to the normal decision region": Check whether the real-time indicator value vector meets the following two conditions simultaneously for most of the observation window (e.g., more than 80% of the data points): (a) it does not exceed any static threshold boundary; (b) for all dynamic association boundaries, its calculated effective deviation Deff is zero (or less than a minimum threshold), that is, it is within the normal association space defined by all dynamic association boundaries.

[0113] "and remain stable": This means that within the observation window, the statistical characteristics of the indicator value vector (such as the moving mean and variance) no longer show a trend of continuous deterioration, or fluctuate around a normal level.

[0114] If both of the above conditions are met, the event sequence (including: the indicator value vector at the time of triggering, the dynamic correlation boundary identifier of the trigger, the control action executed, and the effect data within the observation window) is marked as "effective intervention". Otherwise, if the indicator value fails to return to the normal range within the window, or deteriorates rapidly again after returning, it is marked as "intervention to be optimized".

[0115] Data containing event sequences and intervention effect labels are used as training samples to adjust parameters. The specific implementation method is as follows, which is essentially a reinforcement learning process based on policy evaluation:

[0116] Optimization objective: To enable the system's decision-making strategy (defined by the parameter set {δ, W, T}, where δ is the confidence interval width of a certain boundary, W is its severity weight, and T is the risk threshold of a certain control action level associated with it) to maximize the long-term cumulative "effective intervention" reward, while minimizing false positives and false negatives.

[0117] State and Action: Each decision-making intervention event is viewed as a "state-action-reward" sequence. The state is the system context that triggers the decision (which can be simplified to a risk value R and boundary markers); the action is the level of control action performed; the reward is assigned based on whether it is an "effective intervention" or an "intervention requiring optimization," for example, "effective intervention" receives a positive reward (+1), and "intervention requiring optimization" receives a negative reward (-1). If a more serious anomaly occurs before the preset observation window is reached after the intervention, a larger negative reward is given.

[0118] Parameter adjustment strategy:

[0119] For the confidence interval width δ of the dynamic association boundary: If a boundary frequently triggers "intervention to be optimized," it is judged as overly sensitive, leading to unnecessary intervention. In subsequent iterations, its δ value can be appropriately increased to widen the boundary and allow it to accommodate more normal fluctuations. Conversely, if a fault mode associated with a boundary is missed, and a fault is subsequently confirmed but did not trigger the boundary, its δ value can be reduced to tighten the boundary.

[0120] Regarding the severity weight W of the failure mode: If the calculated risk value R is significantly high (leading to overreaction) for events marked as "needing optimization intervention", the weight W of this trigger boundary can be appropriately reduced. Conversely, for events that are "effectively intervened" but have a low risk value R, almost at the threshold, the weight W can be appropriately increased to ensure that similar situations in the future can trigger corresponding actions more reliably.

[0121] Regarding the overall risk threshold T for control action levels: If a control action at a certain level (such as "process parameter fine-tuning") is repeatedly marked as "intervention requiring optimization" (insufficient or excessive response), the system can automatically fine-tune the risk threshold for that level. For example, if the "fine-tuning" action frequently fails to prevent deterioration, necessitating an escalation to "pause," its threshold T can be lowered. adjust This allows it to be triggered earlier; conversely, it can increase the threshold.

[0122] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligent quality anomaly detection and decision-making based on the Industrial Internet, characterized in that, Specifically, the following steps are included: S1. Through an intelligent gateway with adaptive protocol parsing, collect real-time data of at least three interrelated quality indicators in the production unit, and generate an initial credibility score Ci(t) for each data point based on multi-dimensional verification during the collection process. This score is attached to the corresponding data point as the initial value of the credibility label. S2. Based on the initial confidence score Ci(t) carried by the confidence label, perform real-time patching and lightweight statistical process control screening on the data stream at the edge side based on time series prediction; for the patched data points, update their confidence label values ​​to the edge confidence score Ce(t); for the data that passes the screening, append an "edge screening passed" flag; finally, output the preprocessed data stream with the updated confidence label and screening flag; S3. Based on the historical preprocessed data stream, construct a composite decision model, which includes a static threshold boundary and at least one dynamic correlation boundary, wherein each dynamic correlation boundary is pre-assigned a boundary type identifier representing a specific fault mode and a fault mode severity weight W. The process of constructing the dynamic association boundary includes: Preprocessed data stream segments that consistently bear the "edge screening passed" label and are marked as normal operating conditions are selected from historical data to form a highly reliable training sample set; The implicit correlation function f(xi, xj, ...) = 0 was obtained by fitting the data using a multivariate nonlinear regression method with L2 regularization. Calculate the residual distribution from the training samples to the function f=0, and define the confidence interval width using the predetermined higher-order quantile of this distribution to form a dynamic correlation boundary; S4. For the preprocessed data stream uploaded in real time, extract its indicator value vector and execute: First-level judgment: Determine whether the indicator value vector exceeds the static threshold boundary; Second-level judgment: If not exceeded, proceed with the following steps: S41. Obtain the real-time edge credibility score corresponding to the key indicators that constitute the nearest dynamic association boundary. S42. Based on the obtained real-time edge credibility score, calculate the effective deviation of the index value vector from the boundary; S43. Based on the effective deviation and the fault mode severity weight W associated with the dynamic correlation boundary, generate a comprehensive risk value R; S5. Predefine multiple control action levels and set a comprehensive risk threshold for each level; based on the comprehensive risk value R output by the real-time decision-making stage and the boundary type identifier of the dynamic associated boundary it triggers, select the control action corresponding to the highest level that its comprehensive risk value R exceeds and execute it. S6. Collect the effect data after the control action is executed, and optimize the confidence interval width of the dynamic association boundary, the fault mode severity weight W, and the comprehensive risk threshold of the control action level through reinforcement learning.

2. The intelligent quality discrimination and decision-making method based on the Industrial Internet according to claim 1, characterized in that, The generation of the initial credibility score Ci(t) in S1 includes: The device communication protocol is automatically identified by the protocol probe, and the corresponding parser is loaded. The system performs instantaneous value jump detection, message continuity verification, and logical consistency verification between numerical values ​​and device status in parallel to obtain a comprehensive verification result. Based on the comprehensive verification results, the initial confidence score Ci(t) is calculated.

3. The intelligent quality discrimination and decision-making method based on the Industrial Internet according to claim 1, characterized in that, The update process of the credibility label in S2 is as follows: For data points in the credibility label whose initial credibility score Ci(t) is lower than the first threshold, a repair value is generated based on their historical sequence using a prediction model. The credibility label value of the successfully repaired data points is updated to Ce(t)=k*Ci(t)+(1-k)*Cp, where Cp is the preset base credibility of the repair operation, Ci(t) is the initial credibility score, and k is the weight coefficient.

4. The intelligent quality discrimination and decision-making method based on the Industrial Internet according to claim 1, characterized in that, The second level of judgment includes: Calculate the shortest geometric distance D from the index value vector X to the nearest dynamic association boundary; Identify the key indicator pairs (xm, xn) that constitute the boundary and obtain their corresponding real-time edge confidence scores Cem and Cen; The effective deviation is calculated using the formula Deff=D×(1+γ / (min(Cem,Cen)+ε)), where γ is the sensitivity coefficient and ε is the decimal point to prevent zeroing. The overall risk value R is calculated using the formula R=W×(Deff / Dref), where Dref is the reference distance constant and W is the severity weight of the failure mode.

5. The intelligent quality discrimination and decision-making method based on the Industrial Internet according to claim 1, characterized in that, The control action levels include at least early warning, process parameter fine-tuning, production step suspension, and overall line shutdown; when executing control actions, the boundary type identifier of the triggered dynamic associated boundary is output synchronously as fault pre-diagnosis information.

6. The intelligent quality discrimination and decision-making method based on the Industrial Internet according to claim 1, characterized in that, The optimization in S6 includes: after each control action is executed, an effect observation window is started to collect subsequent quality indicator data and its credibility label; Analyze the data within the observation window. If the indicator value vector regresses to the normal decision region and remains stable, it is marked as an effective intervention; otherwise, it is an intervention to be optimized. Data containing event sequences and intervention effect labels are used as training samples to adjust the confidence interval width of the dynamic association boundary that triggers the intervention, the failure mode severity weight W, and the comprehensive risk threshold for increasing or decreasing the level of its associated control actions.

Images