A quality prediction system and method based on multi-dimensional feature coupling driving

Abstract

The application relates to the fields of data processing and intelligent prediction, in particular to a quality prediction system and method based on multi-dimensional feature coupling driving, which comprises a business logic solidification module, a real-time working condition comparison module, a risk trend extrapolation module and a final risk grading module; the business logic solidification module is used for refining and solidifying business influence logic and generating static risk factors; the real-time working condition comparison module is used for comparing real-time operation data and generating a dynamic deviation file; the risk trend extrapolation module is used for processing the time sequence of the dynamic deviation file and outputting a forward-looking risk track; and the final risk grading module is used for integrating the static risk factors and the forward-looking risk track, grading and outputting a final quality risk evaluation value. The application solves the problem that traditional methods are difficult to integrate multi-dimensional information by deeply coupling the static inherent risk of a business process with the dynamic evolution trend of real-time data, quantifies the global correlation effect of the current deviation in the business process, and improves the accuracy and interpretability of quality risk prediction.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing and intelligent prediction, in particular to a quality prediction system and method based on multi-dimensional feature coupling driving. BACKGROUND

[0002] In the fields of modern industrial production, energy management and environmental monitoring, predicting future states and assessing potential risks by analyzing massive operation data has become a core requirement to ensure system safety and improve operational efficiency. For example, in manufacturing, accurate prediction of product quality risks can guide dynamic adjustment of the production process; in power grid dispatching, risk assessment of power quality is crucial to maintaining grid stability.

[0003] To achieve this goal, existing technologies mostly rely on advanced data-driven paradigms, particularly deep learning-based models. These methods build complex neural network structures, such as using graph neural networks to capture spatial relationships between entities, or using long short-term memory networks and temporal convolution networks to learn the temporal dependencies of historical data, and directly predict future operational indicators.

[0004] However, existing technologies generally have difficulties in integrating business knowledge, lack of bias modeling ideas, and rough integration of static and dynamic risks, which restricts the performance of prediction models in accuracy, stability and explainability.

[0005] Therefore, a quality prediction system and method based on multi-dimensional feature coupling driving are proposed. SUMMARY

[0006] The purpose of the present application is to provide a quality prediction system and method based on multi-dimensional feature coupling driving, aiming to solve the problem of one-sided and lack of foresight in existing quality prediction results. To solve the problems existing in the prior art, the present application extracts and solidifies the business impact logic through a business logic solidification module to generate static risk factors; a real-time operating condition comparison module compares real-time operation data to generate a dynamic deviation archive; a risk trend extrapolation module processes the time series of the dynamic deviation archive to output a forward-looking risk trajectory; and a final risk grading module integrates the static risk factors and the forward-looking risk trajectory to grade and output the final quality risk assessment value.

[0007] To achieve the above purpose, the present application provides a quality prediction system based on multi-dimensional feature coupling driving, which specifically includes the following modules:

[0008] Business logic solidification module: used to collect historical operation data of the monitored process, and extract and solidify the inherent impact logic between business entities according to the pre-set business knowledge ontology to form a structured business rule library; processing the structured business rule library to generate static risk factors;

[0009] a real-time condition comparison module configured to compare the real-time operation data of the monitored process with the structured business rule base as a benchmark to determine the current state deviation between business entities and generate a dynamic deviation profile at each time step;

[0010] a risk trend extrapolation module configured to process the time series of the dynamic deviation profile, induce the deviation information at each time step through an associated effect processor, and generate a phased state summary; input the phased state summary into a time series extrapolator, extrapolate the future trend of the state deviation according to a preset evolution algorithm, and output a forward-looking risk trajectory;

[0011] a final risk grading module configured to integrate the forward-looking risk trajectory and the static risk factor to form a comprehensive input; apply the comprehensive input to a preset scoring logic to grade and output a final quality risk assessment value.

[0012] Preferably, the step of generating the static risk factor specifically comprises: representing the business knowledge ontology as a process topology graph containing nodes and directed edges, wherein the nodes represent business entities and the directed edges represent the influence transmission paths between entities; defining the inherent basic severity level for each node in the process topology graph and the influence propagation coefficient for each directed edge by a knowledge engineering method, as the core content of the structured business rule base; extracting the basic severity levels of all business entities in the structured business rule base to form the static risk factor.

[0013] Preferably, the step of generating the dynamic deviation profile specifically comprises: calculating the monitoring indicators corresponding to each business entity from the historical healthy operation data, and storing the statistical benchmark distribution of the monitoring indicators, wherein the statistical benchmark distribution contains the expected mean and the standard deviation; at each time step, obtaining the real-time operation data value, and calculating the standardized discrete score according to the statistical benchmark distribution corresponding to the real-time operation data value; taking the standardized discrete score as the quantitative current state deviation, recording the standardized discrete scores of all business entities to form the dynamic deviation profile.

[0014] Preferably, the step of generating the periodic state summary specifically comprises: loading a process topology; for each node in the process topology, setting the current state deviation of the node in the dynamic deviation profile as the initial deviation value of the node; iteratively updating the deviation value of each node based on the deviation values of the nodes in the previous iteration according to the influence transmission path of the topology until the iteration is terminated after any condition is met: one, in the adjacent two iterations, the absolute value of the difference between the updated deviation value and the deviation value before updating of all nodes is calculated, and the maximum value of the absolute value is less than a preset convergence threshold; two, a preset maximum iteration number limit is reached; the final deviation values of the nodes are taken as the cumulative deviation vector representing the global correlation effect, i.e. the periodic state summary.

[0015] Preferably, the step of outputting the prospective risk trajectory specifically comprises: the time series extrapolator is a state space predictor; the state space predictor internally maintains a state vector representing the risk state, which includes the current level of risk and the change rate of risk; when processing the time series of the periodic state summary, the state space predictor alternately performs time update and measurement update at each time step, through the time update step, the risk state at the current time is predicted according to the state vector at the previous time, and a prediction result is generated; through the measurement update step, the prediction result is corrected combined with the actual observation value at the current time to obtain the optimal estimate of the risk state at the current time; after the entire time series processing is completed, the state space predictor continuously applies the time update step to extrapolate multiple steps into the future according to the finally obtained state vector, and generates a state vector sequence containing the prediction values of multiple time steps in the future as the prospective risk trajectory.

[0016] Preferably, the step of constructing the comprehensive input specifically comprises: for each business entity, extracting the basic severity level and the risk prediction peak value of the business entity from the static risk factor and the prospective risk trajectory, respectively; taking the extracted basic severity level and risk prediction peak value as the input variables associated with each other, and applying them to a predefined risk assessment model to output a quantitative comprehensive risk index; the comprehensive risk indices generated by all business entities are collected into a set to constitute the comprehensive input; the risk assessment model embodies the nonlinear interaction between the basic severity level and the risk prediction peak value.

[0017] Preferably, the step of grading and outputting the final quality risk assessment value specifically comprises: the preset scoring logic is a risk assessment matrix, the risk assessment matrix pre-defines a plurality of non-overlapping numerical intervals, and each interval is mapped to a unique and ordered quality risk level; the highest comprehensive risk index is matched with the numerical interval of the risk assessment matrix; and a quality risk level corresponding to the interval where the highest comprehensive risk index is located is output as the final quality risk assessment value.

[0018] Preferably, a quality prediction method based on multi-dimensional feature coupling driving comprises:

[0019] Historical operation data of a monitored process are collected, and inherent influence logic between business entities is refined and solidified according to a preset business knowledge ontology to form a structured business rule base; the structured business rule base is processed to generate static risk factors;

[0020] At each time step, the structured business rule base is taken as a benchmark to compare real-time operation data of the monitored process to determine a current state deviation between business entities and generate a dynamic deviation archive;

[0021] The time sequence of the dynamic deviation archive is processed, deviation information at each time step is induced by an associated effect processor to generate a phased state summary; and the phased state summary is input into a time sequence extrapolator to extrapolate a future trend of the state deviation according to a preset evolution algorithm and output a forward-looking risk trajectory;

[0022] The forward-looking risk trajectory and the static risk factors are integrated to form comprehensive input; and the comprehensive input is applied to a preset scoring logic to grade and output a final quality risk assessment value.

[0023] Compared with the prior art, the present application has the following beneficial effects:

[0024] 1. The present application first solidifies the internal logic of a business process and expert knowledge into static risk factors, so that the prediction model is established on the basis of stable business rules rather than completely relying on fluctuating real-time data. This makes the final risk assessment conclusion have clear business logic as support, overcomes the defect that a pure data model is difficult to explain its internal decision-making process, and makes the prediction result more persuasive.

[0025] 2. By simulating the propagation and accumulation of deviations in a business process network through an associated effect processor, the present application can quantify the chain effect of a single local anomaly on the whole. This overcomes the limitation of traditional methods that analyze each monitoring point in isolation, can effectively identify potential systemic risks coupled by multiple minor deviations, and significantly improves the accuracy and integrity of risk identification.

[0026] 3、The application uses a time sequence extrapolator to predict the evolution trend of risks, and generates future risk trajectories. This enables the system to change from passive response alarm to active preventive early warning, and helps managers to take intervention measures to prevent quality problems from the source and reduce potential losses.

[0027] 4、The application integrates static risks reflecting the inherent importance of business entities with predicted risks reflecting their future dynamic changes. This multi-dimensional comprehensive consideration mechanism can more reasonably distinguish risks of different natures, such as slight abnormalities in a core link and severe fluctuations in a non-critical link, thereby helping users to prioritize resources to the most critical risk points, making the decision more efficient and more targeted. BRIEF DESCRIPTION OF DRAWINGS

[0028] Figure 1 A quality prediction system structure diagram based on multi-dimensional feature coupling driving according to the application;

[0029] Figure 2 A quality prediction system and method based on multi-dimensional feature coupling driving according to the application overall interaction timing diagram;

[0030] Figure 3 A multi-dimensional feature processing and risk fusion flowchart according to the application. DETAILED DESCRIPTION

[0031] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative labor fall within the scope of the application.

[0032] Please refer to Figures 1 to 3 The application provides a quality prediction system and method based on multi-dimensional feature coupling driving, and the technical solutions are as follows:

[0033] A quality prediction system based on multi-dimensional feature coupling driving, comprising:

[0034] A business logic solidification module: used for collecting historical operation data of a monitored process, and extracting and solidifying inherent influence logic between business entities according to a preset business knowledge ontology, to form a structured business rule library; processing the structured business rule library to generate static risk factors;

[0035] Real-time working condition comparison module: used for comparing real-time operation data of the monitored process with the structured business rule base as a benchmark to determine current state deviation between business entities and generate a dynamic deviation profile at each time step;

[0036] Risk trend extrapolation module: used for processing time series of the dynamic deviation profile, inducing deviation information at each time step through an associated effect processor to generate a phased state summary; inputting the phased state summary into a time series extrapolator to extrapolate future trend of the state deviation according to a preset evolution algorithm and output a forward-looking risk trajectory;

[0037] Final risk grading module: used for integrating the forward-looking risk trajectory and the static risk factor to form a comprehensive input; applying the comprehensive input to a preset scoring logic to grade and output a final quality risk assessment value.

[0038] Embodiment one

[0039] This embodiment provides a specific application of a quality prediction system and method based on multi-dimensional feature coupling driving, and takes a key thin film deposition process in semiconductor manufacturing as an example to illustrate the specific implementation of the present application. Referring to Figure 2 , Figure 3 , this embodiment detects according to the flow of the quality prediction method based on multi-dimensional feature coupling driving provided by the present application, and shows the data flow.

[0040] Referring to Figure 1 , the system structure of the present application is shown in the figure.

[0041] Further, historical operation data of the monitored process is collected, and inherent influence logic between business entities is refined and solidified according to a preset business knowledge ontology to form a structured business rule base; the structured business rule base is processed to generate a static risk factor, and the specific steps are as follows:

[0042] The system first executes the function of the business logic solidification module in the initial deployment stage. According to the domain knowledge of experienced process engineers, the system constructs a business knowledge ontology describing the deposition process. The business knowledge ontology is represented as a flow topology graph containing nodes and directed edges, wherein the nodes respectively represent gas flow controllers, chamber temperature controllers, and film thickness as the final quality characterization, and the directed edges clearly depict the influence transmission paths between them. Through the knowledge engineering method, the flow topology graph is endowed with quantitative logical parameters, and each node representing a process parameter is defined with an inherent basic severity level. The level of the chamber temperature controller is set to 5 (the highest level is 5), because it is crucial to the final quality; at the same time, each influence path is assigned an influence transmission coefficient, such as the coefficient from the chamber temperature to the film thickness is set to 0.8 to represent the transmission strength of the influence. These parameters together constitute a structured business rule base, and the extracted basic severity level set forms a static risk factor.

[0043] The knowledge engineering method further includes an offline calibration step of the influence transmission coefficient and the basic severity level. The system periodically extracts a dynamic deviation profile and corresponding final product quality data within a period of time from historical data records; using a preset statistical correlation analysis method, the actual correlation strength between the state deviation of each business entity and the final product quality within the period is calculated and taken as a data-driven correlation degree; the data-driven correlation degree is compared with the influence transmission coefficient of the corresponding directed edge in the flow topology graph, and the influence transmission coefficient is adjusted according to the difference between the two; the basic severity level of the specific business entity is adjusted according to the frequency and severity of quality problems caused by the deviation of the specific business entity within the period.

[0044] The offline calibration step combines the initial expert knowledge with long-term actual production data, enhancing the accuracy and reliability of the entire prediction system in long-period operation, and avoiding the decline in prediction performance caused by outdated knowledge base.

[0045] The present application converts domain knowledge into a structured flow topology graph containing basic severity levels and influence transmission coefficients. This step provides clear, business logic-based parameters for subsequent risk calculation, making the evaluation process traceable and interpretable. This approach reduces the dependence on a large number of labeled samples in pure data-driven models and ensures that the evaluation results are consistent with the actual process logic.

[0046] Further, at each time step, the structured business rule base is used as a reference to compare the real-time operation data of the monitored process to determine the current state deviation between business entities and generate a dynamic deviation profile, with the specific steps being as follows:

[0047] After the system is put into operation, the real-time working condition comparison module continues to work. The system calculates the monitoring indicators corresponding to each business entity from a large amount of historical health production batch data in advance, and stores the statistical baseline distribution thereof. A data window with a preset fixed length is defined; the data window is used to select a latest data subset from the historical health operation data; the calculation of the statistical baseline distribution is performed only based on the data subset contained in the data window; as new health operation data is generated, the data window slides forward to include new data and exclude the oldest data, and the statistical baseline distribution is periodically recalculated and updated accordingly. For the monitoring indicator of chamber temperature, the statistical baseline distribution is determined as an expected mean value of 200°C and a standard deviation of 0.5°C. At any production time step, when the real-time operation data value of the chamber temperature obtained by the module is 201.2°C, it will immediately calculate the standardized discrete score of 2.4 corresponding to the value according to the statistical baseline distribution. The score and the standardized discrete scores of other entities (such as gas flow controllers) together constitute the dynamic deviation record at that moment, accurately quantifying the degree of deviation of the current working condition from the ideal state.

[0048] By limiting the calculation of the statistical baseline distribution to the latest data subset, the present application ensures that the baseline can reflect the actual running state of the monitored process in the near future, improves the accuracy of the subsequently calculated standardized discrete scores, and thus provides more timely and reliable input data for the entire risk prediction system.

[0049] The present application uses a method based on the statistical distribution of historical data to calculate the deviation degree of the current state. This method converts different types of monitoring data into standardized, dimensionless values, facilitating unified subsequent processing and comparison. It takes into account the historical fluctuation range of the data, providing a relatively objective baseline for determining the current deviation, which helps to reduce false positives that may occur due to improper fixed threshold settings.

[0050] Further, the time series of the dynamic deviation record is processed, and the deviation information at each time step is summarized by an association effect processor to generate a stage summary; the stage summary is input into a time series extrapolator, which extrapolates the future trend of the state deviation according to a preset evolution algorithm, and outputs a forward-looking risk trajectory, the specific steps being as follows:

[0051] Subsequently, the risk trend extrapolation module receives and processes the time series composed of the dynamic deviation record. First, the association effect processor loads the process topology graph, and for each node in the graph, sets the current state deviation of the node in the dynamic deviation record as the initial deviation value of the node. According to the influence transmission path of the topology graph, the processor starts to iteratively update the deviation value of each node.

[0052] In calculating the contribution of each upstream node to its downstream nodes, the bias value of the upstream node is first processed by a preset nonlinear activation function; the nonlinear activation function has a saturation characteristic, which is that when the absolute value of the input bias value is small, the output is approximately proportional to the input, and when the absolute value of the input bias value exceeds a certain range, the output tends to a constant upper limit value; the output value after the nonlinear activation function processing is multiplied by the influence propagation coefficient, as the effective influence input of the upstream node to the downstream node.

[0053] By nonlinearly transforming the bias value of the upstream node, the system can effectively limit the excessive influence of extreme abnormal values of a single node on the cumulative bias evaluation of the entire system. This step enhances the calculation stability of the correlation effect processor when facing extreme abnormalities or noise in the input data, making the results of the periodic state summary more stable and reasonable, and providing higher-quality input data for subsequent trend extrapolation.

[0054] In the adjacent two iterations, the processor calculates the absolute value of the difference between the updated bias value and the bias value before updating of all nodes, and finds the maximum value. When the maximum value is less than the preset convergence threshold 0.01, the iteration is terminated. The final obtained bias value of each node constitutes a cumulative bias vector representing the global correlation effect, that is, the periodic state summary at this moment.

[0055] The present application simulates the mutual influence between different business entities by iterative calculation on the process topology graph. This method aims to quantify the cumulative effect that may be generated by the propagation of local bias in the system, so as to obtain a comprehensive index reflecting the overall state at the current moment. This helps to identify potential associated risks formed by the combined action of multiple independent and smaller biases.

[0056] Next, the state space predictor as the time series extrapolator starts to work. It maintains a state vector containing the current level and change rate of risk. In processing the time series of periodic state summary, the predictor alternately performs time update and measurement update at each time step: first, through the time update step, the risk state at the current moment is predicted according to the state vector at the previous moment; then, through the measurement update step, the prediction result is corrected by combining the actual observation value at the current moment to obtain the optimal estimate of the current risk state. When the entire historical time series is processed, the predictor continues to apply the time update step to extrapolate multiple steps into the future based on the finally obtained state vector, generating a state vector sequence containing the prediction values of multiple future time steps, as the prospective risk trajectory. For example, the trajectory may show that the comprehensive risk will reach a peak of 4.5 at the third time step in the future.

[0057] The application adopts a state space predictor to process time series data representing overall states. The predictor can estimate the future change trend of the historical sequence according to the dynamic characteristics of the sequence, and generate a forward-looking data track. The track provides time dimension input for subsequent risk assessment, so that the assessment can take into account the future state.

[0058] Further, the forward-looking risk track and the static risk factor are integrated to form a comprehensive input; the comprehensive input is applied to a preset scoring logic to grade and output a final quality risk assessment value, and the specific steps are as follows:

[0059] Finally, the final risk grading module integrates and judges the dynamic and static information. For the chamber temperature controller entity as a risk source, the module extracts its basic severity level 5 and risk prediction peak 4.5 from the static risk factor and the forward-looking risk track respectively. The two values are used as interrelated input variables and applied to a predefined risk assessment model.

[0060] When extracting the risk prediction peak from the forward-looking risk track, the future time step corresponding to the risk prediction peak is also extracted; according to the future time step and a preset time decay coefficient, a time decay weight is calculated; the risk prediction peak and the time decay weight are operated to generate a time-corrected risk prediction peak; the time-corrected risk prediction value is used subsequently to replace the original risk prediction peak to calculate the comprehensive risk index.

[0061] The risk assessment model reflects the nonlinear interaction between the two, when the basic severity level is very high, even if the risk prediction peak is only moderate, the output comprehensive risk index may be amplified. In this example, the quantitative comprehensive risk index output by the model is 28.0. The comprehensive risk indexes generated by all business entities are collected into a set to form a comprehensive input.

[0062] The time decay processing of the risk prediction peak makes the final risk assessment result not only reflect the severity of the risk, but also indirectly reflect the urgency of the risk, improving the guiding value of the early warning information for production intervention.

[0063] The application integrates the static factor representing inherent importance and the dynamic prediction value representing future trend through a model reflecting nonlinear interaction. This processing method attempts to reflect the inherent severity of the event and the possibility of future development in a single numerical value, so that the finally generated comprehensive risk index can contain more information dimensions.

[0064] In the rating step, the system matches the highest comprehensive risk index 28.0 in the input set with a pre-defined risk assessment matrix. The risk assessment matrix has pre-defined a plurality of non-overlapping numerical intervals and corresponding unique quality risk levels. The matrix defines 15-25 as "Level 3 (High Risk)" and greater than 25 as "Level 4 (Extremely High Risk)". Since 28.0 falls in the interval of "Level 4 (Extremely High Risk)", the system outputs this final quality risk assessment value.

[0065] The present application uses a risk assessment matrix to map the calculated comprehensive risk index to a pre-defined risk level. This step converts the continuous numerical result into discrete, ordered level categories, making the assessment output easier to understand and use, facilitating correspondence with existing operating procedures or risk management standards.

[0066] The present application provides a quality risk prediction method. This method integrates static business logic and dynamic real-time data, and extrapolates the future trend of the data to generate a quality risk assessment value for the monitored process. This allows users to obtain a quantitative reference on potential risks before the final product quality is formed, providing data support for production process adjustment, and helping to reduce potential losses caused by unknown risks.

[0067] Example Two

[0068] This example describes the specific workflow of the business logic solidification module and the real-time working condition comparison module in the present application. The scenario is set to the plasma etching process in semiconductor manufacturing, and the final etching uniformity quality of this process is affected by multiple process parameters such as radio frequency power and reaction gas flow.

[0069] In the initial deployment stage, the system first performs the function of the business logic solidification module. According to the knowledge of process experts, the system represents the business knowledge ontology of this etching process as a flow topology graph containing nodes and directed edges. Among them, the nodes represent business entities, such as radio frequency power controller and gas flow controller; the directed edges represent their influence transmission paths on the final product quality. Through knowledge engineering methods, this topology graph is given quantitative logic parameters: define the inherent basic severity level of the radio frequency power controller node as 4 (the highest is 5 levels), and define the basic severity level of the gas flow controller node as 3. At the same time, define the influence propagation coefficient of the edge from the radio frequency power controller to the final quality as 0.9, and define the influence propagation coefficient of the edge from the gas flow controller to the final quality as 0.7. These parameters together constitute a structured business rule base. Subsequently, the system extracts the basic severity levels of all business entities to form static risk factors.

[0070] During the process operation, the real-time condition comparison module is continuously working. The system has calculated and stored the statistical baseline distribution of the monitoring indicator, radio frequency power, from historical healthy production data, with an expected mean value of 500 W and a standard deviation of 2 W. At a certain production time step, the system obtains the real-time operation data value of radio frequency power as 505 W. According to the statistical baseline distribution corresponding to the value, the system calculates the standardized discrete score of 2.5. By performing the same operation on all relevant business entities, the system records the standardized discrete scores of all entities at the time, forming a dynamic deviation archive of the time step. This dynamic deviation archive, together with the archives generated at other time steps, constitutes a time series, which serves as the input data of the subsequent risk trend extrapolation module.

[0071] Embodiment Three

[0072] This embodiment describes the specific workflow of the risk trend extrapolation module and the final risk grading module in the present application. In this embodiment, the system has generated a time series of dynamic deviation archives of the chemical mechanical polishing process through the previous modules and is ready for in-depth trend analysis and risk grading.

[0073] After receiving the time series of dynamic deviation archives, the risk trend extrapolation module begins analysis. First, the correlation effect processor loads the process topology graph at each time step. For a certain time step, it sets the current state deviation in the dynamic deviation archive at that time as the initial deviation value of the corresponding node in the process topology graph. According to the influence transmission path and influence propagation coefficient of the topology graph, the processor updates the deviation value of each node through iterative calculation to simulate the propagation and accumulation of risk in the system. When the maximum absolute value of the difference between the updated deviation value and the previous deviation value of all nodes after each adjacent iteration is less than the preset convergence threshold, the iteration terminates. The deviation values of each node obtained after the iteration constitute a cumulative deviation vector representing the global correlation effect, i.e., the stage summary at that time.

[0074] Next, the state space predictor, as the time series extrapolator, begins processing the time series composed of multiple stage summaries. It maintains a state vector containing the current level and rate of change of risk. When processing the time series, the predictor alternately performs time update and measurement update steps at each time step to accurately track the dynamic evolution of the risk state. After the entire time series is processed, the predictor continuously applies the time update step to the future for multi-step extrapolation based on the final state vector obtained, generating a state vector sequence containing the predicted values of multiple future time steps, i.e., the prospective risk trajectory. Assuming that the trajectory shows that the integrated risk of the system will reach a risk prediction peak value of 5.0 at a certain time in the future.

[0075] Finally, the final risk rating module starts working. It extracts the base severity rating 4 and the risk prediction peak 5.0 from the static risk factor and the prospective risk trajectory, respectively, for the critical business entity of the radio frequency power controller. These two values are applied as interdependent input variables to a predefined risk assessment model that embodies a non-linear interaction. After the model calculation, the output is a quantitative overall risk index of 24. Subsequently, the system collects all overall risk indices generated by the business entities into a set and finds the highest index value of 24 from the set. Finally, a pre-defined rating logic, i.e. a risk assessment matrix, matches this highest overall risk index with pre-defined numerical intervals in the matrix. The matrix defines that the interval of 15-25 corresponds to "level 3 (high risk)", and the system outputs the final quality risk assessment value as "level 3 (high risk)".

[0076] While embodiments of the present application have been shown and described with reference to particular embodiments thereof, it will be understood by those skilled in the art that various changes in form and details can be made therein without departing from the spirit and scope of the application. The scope of the application is defined by the appended claims and their equivalents.

Claims

1. A quality prediction system based on multidimensional feature coupling, characterized in that, include: Business Logic Solidification Module: Used to collect historical operational data of the monitored process, and based on the preset business knowledge ontology, extract and solidify the inherent influence logic between business entities to form a structured business rule base; process the structured business rule base to generate static risk factors; represent the business knowledge ontology as a process topology graph containing nodes and directed edges, where nodes represent business entities and directed edges represent the influence transmission path between entities; Using knowledge engineering methods, an inherent basic severity level is defined for each node in the process topology diagram, and an influence propagation coefficient is defined for each directed edge, serving as the core content of the structured business rule base; the basic severity levels of all business entities in the structured business rule base are extracted to form static risk factors; Periodically extract dynamic deviation files and corresponding final product quality data for a period of time from historical data records; Using a pre-defined statistical correlation analysis method, the actual correlation strength between the state deviation of each business entity and the final product quality within a period is calculated, and this is used as the data-driven correlation degree. The data-driven correlation is compared with the influence propagation coefficient of the corresponding directed edge in the process topology diagram, and the influence propagation coefficient is adjusted based on the difference between the two; the basic severity level of the business entity is adjusted based on the frequency and severity of quality problems caused by deviations of specific business entities within a period. Real-time operational status comparison module: At each time step, the module uses a structured business rule base as a benchmark to compare the real-time operational data of the monitored process, determine the current state deviation between business entities, and generate a dynamic deviation profile. Risk Trend Extrapolation Module: Used to process the time series of dynamic deviation archives, it summarizes the deviation information of each time step through the correlation effect processor and generates a phased state summary; the phased state summary is input into the time series extrapolator, and the future trend of the state deviation is extrapolated according to the preset evolution algorithm, and a forward-looking risk trajectory is output. Final risk rating module: used to integrate forward-looking risk trajectories and static risk factors to form a comprehensive input; The comprehensive input is applied to the preset scoring logic to determine the level and output the final quality risk assessment value.

2. The quality prediction system based on multidimensional feature coupling driven according to claim 1, characterized in that, The steps for generating the dynamic deviation profile specifically include: calculating the monitoring indicators for each business entity from historical health operation data, and storing the statistical baseline distribution of the monitoring indicators, wherein the statistical baseline distribution includes the expected mean and standard deviation; at each time step, acquiring real-time operation data values, and calculating standardized discrete scores based on the statistical baseline distribution corresponding to the real-time operation data values; using the standardized discrete scores as quantified current state deviations, recording the standardized discrete scores of all business entities, and forming the dynamic deviation profile.

3. The quality prediction system based on multidimensional feature coupling driven according to claim 1, characterized in that, The steps for generating the phased state summary specifically include: Load the process topology diagram; for each node in the process topology diagram, set the current state deviation of the node in the dynamic deviation file as the initial deviation value of the node; according to the influence transmission path of the topology diagram, calculate based on the deviation value of each node in the previous iteration, iteratively update the deviation value of each node until the iteration terminates when either condition is met: first, in two adjacent iterations, calculate the absolute value of the difference between the updated deviation value and the original deviation value of all nodes, the maximum value of the absolute value is less than a preset convergence threshold; second, reach the preset maximum number of iterations limit; use the final obtained deviation value of each node as the cumulative deviation vector representing the global correlation effect, i.e., the stage state summary.

4. The quality prediction system based on multidimensional feature coupling driven according to claim 1, characterized in that, The steps for outputting the forward-looking risk trajectory specifically include: the time-series extrapolator is a state-space predictor; the state-space predictor internally maintains a state vector representing the risk state, which includes the current level of risk and the rate of change of risk; when processing the time series of the phased state summary, the state-space predictor alternately performs time update and measurement update at each time step. Through the time update step, the risk state at the current time is predicted based on the state vector at the previous time step; through the measurement update step, the prediction result is corrected by combining the actual observation value at the current time step to obtain the optimal estimate of the risk state at the current time step; after the entire time series is processed, the state-space predictor continues to apply the time update step, extrapolating to the future in multiple steps based on the finally obtained state vector, generating a state vector sequence containing the predicted values ​​of multiple future time steps, as the forward-looking risk trajectory.

5. A quality prediction system based on multidimensional feature coupling driven according to claim 1, characterized in that, The steps for constructing the comprehensive input specifically include: for each business entity, extracting the basic severity level and the predicted risk peak from static risk factors and forward-looking risk trajectories respectively; using the extracted basic severity level and the predicted risk peak as interrelated input variables, applying them to a predefined risk assessment model, and outputting a quantified comprehensive risk index; the comprehensive risk indices generated by all business entities are aggregated into a set to constitute the comprehensive input; the risk assessment model reflects the nonlinear interaction between the basic severity level and the predicted risk peak.

6. A quality prediction system based on multidimensional feature coupling driven according to claim 1, characterized in that, The steps of grading and outputting the final quality risk assessment value specifically include: the preset scoring logic is a risk assessment matrix, which predefines multiple non-overlapping numerical intervals, and each interval is mapped to a unique and ordered quality risk level; matching the highest comprehensive risk index with the numerical interval of the risk assessment matrix; and outputting the quality risk level corresponding to the interval where the highest comprehensive risk index is located as the final quality risk assessment value.

7. A quality prediction method based on multidimensional feature coupling, characterized in that, include: Collect historical operational data of the monitored process, and extract and solidify the inherent influence logic between business entities based on the preset business knowledge ontology to form a structured business rule base; process the structured business rule base to generate static risk factors; represent the business knowledge ontology as a process topology graph containing nodes and directed edges, where nodes represent business entities and directed edges represent the influence transmission path between entities. Using knowledge engineering methods, an inherent basic severity level is defined for each node in the process topology diagram, and an influence propagation coefficient is defined for each directed edge, serving as the core content of the structured business rule base; the basic severity levels of all business entities in the structured business rule base are extracted to form static risk factors; Periodically extract dynamic deviation files and corresponding final product quality data for a period of time from historical data records; Using a pre-defined statistical correlation analysis method, the actual correlation strength between the state deviation of each business entity and the final product quality within a period is calculated, and this is used as the data-driven correlation degree. The data-driven correlation is compared with the influence propagation coefficient of the corresponding directed edge in the process topology diagram, and the influence propagation coefficient is adjusted based on the difference between the two; the basic severity level of the business entity is adjusted based on the frequency and severity of quality problems caused by deviations of specific business entities within a period. At each time step, the structured business rule base is used as a benchmark to compare the real-time operational data of the monitored process, determine the current state deviation between business entities, and generate a dynamic deviation profile. The time series of dynamic deviation archives is processed, and the deviation information of each time step is summarized by the correlation effect processor to generate a phased state summary. The phased state summary is input into the time series extrapolator, and the future trend of the state deviation is extrapolated according to the preset evolution algorithm to output a forward-looking risk trajectory. Integrate forward-looking risk trajectories and static risk factors to form a comprehensive input; apply the comprehensive input to a pre-set scoring logic to determine the level and output the final quality risk assessment value.

Images