An aviation assembly process quality problem occurrence frequency prediction method based on multi-factor risk prediction
By constructing a structured factor library and a dynamic weight fusion model, the problems of data sparsity and high-dimensional coupling of factors in the aviation assembly process are solved, enabling real-time early warning and prediction of quality problems in the aviation assembly process, dynamically adapting to small-batch production, and improving the accuracy and interpretability of predictions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- AVIC SAC COMML AIRCRAFT
- Filing Date
- 2026-04-17
- Publication Date
- 2026-07-21
AI Technical Summary
The aircraft assembly process faces challenges such as data sparsity, high-dimensional coupling of factors, severe defects, difficulty in knowledge integration, and unique prediction targets. Existing quality control methods are difficult to apply effectively, especially in small-batch, multi-variety production scenarios where it is difficult to output the frequency of quality problems occurring within a specific time period.
A structured factor library is constructed, and the relationship between risk factors and defects is defined through a many-to-many mapping matrix. By combining a gradient boosting decision tree model and an induced ordered weighted average operator, expert experience and data-driven weights are dynamically integrated to achieve real-time early warning and prediction of quality problems in the aviation assembly process.
It enables multi-dimensional risk prediction in the aviation assembly process, dynamically adapts to differences in different aircraft models and production lines, provides highly accurate early warning information, facilitates engineers' understanding and intervention, and improves the level of intelligence in the production process.
Smart Images
Figure CN122434249A_ABST
Abstract
Description
Technical Field
[0001] This application belongs to the field of industrial intelligent manufacturing and quality control technology, specifically relating to an online prediction and early warning method for quality in aerospace assembly processes based on data-driven and knowledge fusion. In particular, it relates to a system and method that achieves real-time calculation and closed-loop feedback of quality risks in aerospace assembly processes by constructing a structured factor library, dynamically integrating weights using the induced ordered weighted average (IOWA) operator, and fusing a gradient boosting decision tree (GBDT) model. It is especially suitable for early warning and prediction of the frequency of quality problems in key processes related to aircraft structural assembly and system installation. Background Technology
[0002] The assembly process of aerospace products is characterized by complex structures, extremely high precision requirements, diverse product types, small batch production, and strong coordination relationships. The assembly of a modern aircraft involves tens of thousands of parts and thousands of processes; errors in any stage can lead to serious quality defects or even catastrophic safety accidents. Therefore, quality control in aerospace assembly has always been a research hotspot and a challenge in the industrial manufacturing field.
[0003] Traditional quality control methods primarily rely on post-production inspection and statistical process control (SPC). While these methods have achieved some success in high-volume manufacturing industries such as automotive and electronics, they face the following fundamental challenges in the aerospace assembly field:
[0004] I. The characteristics of small batches and multiple varieties make statistical modeling difficult.
[0005] Industries like automotive and electronics typically feature high-volume, repetitive production, accumulating ample homogeneous data for statistical modeling. However, aerospace assembly is characterized by small-batch, multi-variety, and customized production. The annual output of the same aircraft model is limited, and significant differences exist in the process conditions between different aircraft models and batches. Traditional SPC methods are difficult to apply effectively due to insufficient sample size and unstable data distribution.
[0006] Second, the high degree of coupling among multiple risk factors exceeds the complexity of typical manufacturing industries.
[0007] Quality prediction in general manufacturing often focuses on a few key variables, such as equipment parameters or raw material batches. In contrast, quality risks in aerospace assembly involve six dimensions: personnel, equipment, materials, methods, environment, and testing, with complex interactive couplings between these dimensions. For example, the skill level of operators directly affects the accuracy of equipment operation, while environmental temperature and humidity affect the material compatibility. Existing methods struggle to effectively model such a high-dimensional, nonlinear, and strongly coupled system.
[0008] Third, the severity and irreversibility of the quality defects far exceed those of conventional products.
[0009] Quality defects in products like automobiles and home appliances can usually be remedied through repair or component replacement, with relatively limited impact on safety. However, critical defects in aerospace assembly, such as loose fasteners, out-of-tolerance hole diameters, and seal failures, can lead to reduced structural strength, shortened fatigue life, and even flight accidents. Once these defects enter subsequent processes or are delivered for use, repair costs are extremely high, and may even be irreversible. Therefore, aerospace assembly demands far higher levels of early warning and zero-defect requirements than general manufacturing.
[0010] IV. The difficulty of integrating process knowledge with data
[0011] In industries like semiconductors and pharmaceuticals, process control heavily relies on sensor data and automated closed-loop regulation. However, aerospace assembly relies heavily on the experience and manual operation of skilled workers. Assembly process knowledge, such as tool clearance, torque control, and adhesive application techniques, often exists as tacit knowledge within operators and process documents, making it difficult to directly quantify into sensor parameters. Existing technologies either overemphasize data-driven approaches while neglecting process knowledge, or overemphasize expert rules without dynamic adaptation, lacking a method that can organically integrate data-driven approaches with expert knowledge.
[0012] V. Existing quality prediction methods are unable to output the frequency of problem occurrence within a specific time period.
[0013] Existing technologies largely focus on binary judgments of whether defects occur, or regression predictions of continuous quality indicators such as dimensional deviations. However, in aerospace assembly plants, production managers are more concerned with the frequency of various assembly quality issues occurring within specific timeframes, such as a production run, a workday, or a week, in order to optimize production scheduling, pre-allocate quality inspection resources, and evaluate the effectiveness of process improvements. Currently, there is a lack of intelligent methods that can integrate multi-source data from personnel, equipment, materials, methods, environment, and testing, dynamically quantify the impact of each risk factor, and ultimately output predicted values for the frequency of quality issues.
[0014] In summary, quality prediction in the aerospace assembly process faces multiple challenges, including data sparsity, high-dimensional coupling of factors, severity of defects, difficulty in knowledge fusion, and specific prediction objectives. Conventional manufacturing quality control and prediction methods are difficult to directly transfer and apply. Therefore, there is an urgent need for an intelligent method specifically designed for aerospace assembly scenarios, capable of integrating multi-source data and expert knowledge, dynamically adapting to the characteristics of small-batch production, and outputting predictions of the frequency of quality defects. Summary of the Invention
[0015] This application aims to solve the above problems by proposing an early warning and prediction method for the frequency of quality problems in the aviation assembly process based on multi-factor risk prediction. This method achieves early warning and prediction of quality problems in the assembly process by constructing a factor library, establishing a logical relationship model, and dynamically assessing the risk probability.
[0016] According to one aspect of this application, a method for predicting the frequency of quality problems in aviation assembly processes based on multi-factor risk prediction is provided, comprising the following steps:
[0017] S1 constructs a many-to-many mapping matrix to digitize and computably represent the relationship between factors and defects, including:
[0018] S11 defines a structured factor library, which, based on aviation assembly processes, precisely defines all potential risk factors from six dimensions: personnel (U), equipment (E), materials (M), methods (W), environment (S), and monitoring (I), forming a structured data set that can be identified and processed by computer systems.
[0019] S12 constructs a mathematical matrix Used to store the correlation between risk factors and quality defects, matrix elements Representing the The factor affects the first The initial impact weight of each defect;
[0020] S13 quantifies factor attributes, assigning three quantifiable technical attributes to each factor: frequency of occurrence. Severity Detectability This provides a data foundation for subsequent quantitative calculation of risk values;
[0021] S2 constructs a dynamic risk prediction model that integrates data and knowledge. It effectively merges industrial system data with expert knowledge and outputs quantified risk values through interpretable algorithms, including:
[0022] Data preparation involves collecting historical assembly data, quality records, and process parameters; constructing training and testing sets; and preprocessing the collected data.
[0023] Through standardized data interfaces (API, OPCUA), historical process data and final quality inspection results corresponding to the structured factor library are automatically extracted from the Manufacturing Execution System (MES), Quality Management System (QMS), and shop floor sensor network. ;
[0024] The extracted data undergoes cleaning, alignment, and normalization preprocessing. It is then sliced and labeled according to assembly frame, workstation, and work step, ultimately constructing a structured training dataset suitable for machine learning models. ,in It is a risk factor state vector, which represents the state of each risk factor in this assembly process;
[0025] Structured training dataset The training sets are divided chronologically into sets used for model training, parameter tuning, and model selection. and the test set used for final performance evaluation ;
[0026] This model is the core technology of this application. Its goal is to effectively integrate industrial system data with expert knowledge and output a quantified risk value through an interpretable algorithm.
[0027] The S3 training base probability prediction model uses Gradient Boosting Decision Tree (GBDT) as the base learner. Learn a vector from the risk factor state for input. Probability of occurrence of underlying quality defects mapping function This means learning purely statistical correlations from data; and continuously optimizing parameters to reduce the probability of defects occurring. Approaching the final quality inspection result in an infinitely close manner (True label), that is, in a statistical sense ;
[0028] This model can effectively process tabular data, capture complex nonlinear relationships between high-dimensional factors, and the output of model feature importance helps in subsequent risk tracing analysis.
[0029] S4 introduces a dynamic weight fusion module, which incorporates an Induced Ordered Weighted Average (IOWA) operator. Using the historical prediction accuracy of expert experience weights and data-driven weights as the "inducible value," it dynamically generates the optimal weighted vector. This dynamic weight fusion module receives expert experience weights from the many-to-many mapping matrix in S1. And data-driven weights derived from the feature importance of gradient boosting decision tree-based probabilistic prediction models in S3. The IOWA operator is used for fusion to output the final optimized weights. It overcomes the static and one-sided nature of a single weight source.
[0030] S5 calculates the final risk using a risk value quantification engine that integrates FMEA logic. The core calculation formula is as follows:
[0031]
[0032] Indicates defects The overall risk value, that is, the defect under the current factor state. The overall risk level, with a range of positive real numbers, is used for early warning judgment;
[0033] Representation factor The frequency of occurrence, i.e., the factor The frequency level in the historical assembly process ranges from 1 to 10, with a higher value indicating more frequent occurrence. The data is based on historical data statistics.
[0034] Indicates defects The severity, i.e., the defect The severity level of the impact on product functionality and safety ranges from 1 to 10, with higher values indicating greater severity.
[0035] Representation factor The detectability, that is, the ability of existing detection methods to promptly detect and intercept factors. The difficulty level is categorized into 1 to 10, with higher values indicating greater difficulty in detection.
[0036] This represents the probability of defect occurrence predicted by the GBDT model, i.e., in the factor state vector. The probability of a defect occurring is given below, with a value range of [0,1].
[0037] This indicates the final optimization weight factor for defects. The fusion weights are output by the IOWA operator and have a value range of [0,1].
[0038] For the factor in The state vector at time t, describing The state of all risk factors at any given time, with values ranging from a multi-dimensional vector, is derived from real-time data from the manufacturing execution system, quality management system, and workshop sensor network.
[0039] when When the preset threshold is exceeded, an early warning message will be automatically generated and pushed.
[0040] Technically, it organically integrates the traditional FMEA analysis framework (through... ) and the predictive power of modern machine learning (through (and), which unifies qualitative experience with quantitative data, making the output results both highly accurate and highly interpretable.
[0041] S6 deployment;
[0042] S7 self-learning and closed-loop update.
[0043] The S6 deployment specifically includes:
[0044] The trained dynamic risk prediction model is encapsulated as an independent RESTful API microservice, deployed on the production server, and integrated with the workshop MES system to continuously monitor or periodically pull real-time assembly data streams; for each set of received real-time factor status data... The service completes risk assessment within seconds. The calculation.
[0045] The S7 self-learning and closed-loop update specifically include:
[0046] The system automatically associates the context of each warning with the final quality inspection result label of the production task, forming a structured "warning-result feedback pair". It periodically (e.g., weekly) initiates an automated model update task, using the accumulated feedback pairs as incremental data to retrain the GBDT model and refresh the final optimized weights. ;
[0047] The context includes the risk factor status, predicted value, and timestamp.
[0048] This enables the system to adapt to slow changes in the production process (such as equipment aging and process improvements), achieving a technological leap from a "static model" to a "continuous evolution model".
[0049] Visualized early warning and human-computer interaction front end:
[0050] Technical presentation: Develop a web-based visual dashboard that uses the form of an electronic dashboard in the workshop to graphically display the real-time risk index of each workstation and flight sequence (such as heat maps and trend charts);
[0051] Precise identification: When the risk exceeds the threshold, the system not only alarms, but also uses techniques such as highlighting and pop-up windows to accurately identify one or more key factors that cause the risk (such as "E5_Equipment Failure") and associate them with the specific defect type that may be caused (such as "Please note H1_Aperture Out of Tolerance"), providing engineers with direct and clear technical basis for intervention.
[0052] Beneficial effects of this application
[0053] 1. Early warning: Potential quality risks can be identified in advance during the assembly process to avoid defects from occurring;
[0054] 2. Multi-dimensional integration: Integrating six major factors—people, machines, materials, methods, environment, and measurement—to comprehensively cover the entire aerospace assembly process;
[0055] 3. Dynamic Adaptation: The model parameters are dynamically adjusted through machine learning algorithms to adapt to the differences between different machine models and production lines;
[0056] 4. High interpretability: The mapping between factors and problems is clear, making it easy for engineers to understand and intervene;
[0057] 5. Strong compatibility: It can be integrated with existing MES and QMS systems to improve the level of intelligent manufacturing. Attached Figure Description
[0058] Figure 1 This is a schematic diagram of the overall process of the prediction method.
[0059] Figure 2 This is a schematic diagram of the factor-defect mapping matrix.
[0060] Figure 3 This is a flowchart of dynamic weight integration based on the IOWA operator. Detailed Implementation
[0061] The present application is described in detail below with reference to the embodiments, but the present application is not limited to these embodiments.
[0062] Example 1
[0063] The specific implementation methods of this application will be described in detail below with reference to the accompanying drawings and technical solutions. The core of this method lies in systematizing and digitizing the scattered factors of "people, machines, materials, methods, environment, and measurement" in aerospace assembly, and establishing a dynamic prediction model between them and the final quality defects.
[0064] The specific steps are as follows:
[0065] Step 1: Constructing a multi-dimensional factor library and mapping relationships.
[0066] Systematic definition of factors: Based on the characteristics of aviation assembly processes, all potential risk factors (such as "P2_not operating according to document requirements", "E5_equipment failure", "M7_mismixing of materials") are precisely defined from six dimensions: personnel (P), equipment (E), materials (M), methods (W), environment (S) and monitoring (I), forming a structured factor library.
[0067] Establish a defect mapping matrix: Based on historical failure data and expert experience, establish a matrix of many mapping relationships between factors and specific assembly or final product quality problems (such as "H1_hole diameter out of tolerance", "R10_fastener loose", "D2_dent damage"). This matrix not only records correlations, but more importantly, it assigns an initial influence weight to each pair of "factor-defect" relationships. .
[0068] Factor attribute quantification: Assigning three key attribute values to each factor:
[0069] Frequency of occurrence Frequency of occurrence of this factor based on historical data statistics
[0070] Severity This factor leads to defects. The severity level of this defect on product functionality and safety (typically 1-10).
[0071] Detectability The probability that existing control methods can detect and intercept this factor in a timely manner.
[0072] Step 2: Data fusion and preprocessing during assembly.
[0073] Multi-source data acquisition: Extracting historical process data and final quality inspection results corresponding to the factor library from MES (Manufacturing Execution System), QMS (Quality Management System), process documents, and sensors. ).
[0074] Data alignment and labeling: The process data is aligned according to assembly stand, workstation, and work step, and the final inspection result is used as the label to construct a labeled dataset. in It is a multidimensional vector that represents the state of each factor during the assembly process.
[0075] Dataset partitioning: dividing the dataset Divided into training sets according to time sequence (Used for model training, parameter tuning, and model selection), test set (For final performance evaluation).
[0076] Step 3: Constructing a risk prediction model based on weight optimization.
[0077] Model Selection and Initialization: A machine learning model with good interpretability and the ability to handle high-dimensional data (such as Gradient Boosting Decision Tree (GBDT)) was selected as the base model. The goal of the model is to learn a decision tree from factor states. To the probability of defect occurrence Mapping function: Meanwhile, continuously optimize parameters to reduce the probability of defects occurring. Approaching the final quality inspection result in an infinitely close manner (True label), that is, in a statistical sense .
[0078] Dynamic Weighted Integration: Introducing the Induced Ordered Weighted Average (IOWA) operator to dynamically integrate expert experience and data-driven results. For the... The factor in the th The weight of each defect, its final weight Calculated by the following formula:
[0079]
[0080] in, As the initial expert experience weight, The weights that the model learns from the data. This is a dynamic weighted vector related to prediction accuracy.
[0081] Indicates the final optimized weights, factors defects The weights after fusion, with values ranging from [0,1], are derived from the output of the IOWA operator and are used for risk value calculation.
[0082] Represents the dynamic weighting coefficient, the first The fusion coefficients for each weight source are determined based on historical prediction accuracy, and their values range from [0,1]. It is calculated based on the historical prediction accuracy;
[0083] No. The weights of the sources are sorted by induced value (prediction accuracy). Each weight value, ranging from [0,1], is derived from expert experience or data-driven approaches and is relabeled after sorting.
[0084] This represents the weight source index, distinguishing different weight sources. Its value range is {1,2}, and it is fixed at two sources.
[0085] Represents the factor index, the first in the structured factor library. Each risk factor takes a positive integer value and is derived from the factor library definition.
[0086] This represents the defect index, the first in the quality defect type. This type of defect has a value range of positive integers and is derived from the defect database definition.
[0087] Risk value calculation: For factor status data collected in real time Its defects risk value The calculation is as follows:
[0088]
[0089] This formula combines the logic of traditional risk analysis (such as FMEA) with data-driven forecasting results.
[0090] Step 4: Real-time risk prediction and visual early warning.
[0091] Model Deployment: Deploy the trained risk prediction model as an independent prediction service, integrate it with the workshop MES system, and receive real-time data streams.
[0092] Risk Cockpit: Develop a visual early warning interface, such as graphically displaying the real-time risk index of each workstation and each flight. The system not only displays the overall risk value, but also locates specific high-risk factors (such as "Current E5_Equipment Failure Factor Risk Surge") and prompts the possible defect types ("Please note H1_Aperture Out-of-Tolerance Risk").
[0093] Threshold warning: Set dynamic risk thresholds for different levels of defects. .when When this happens, the system will automatically trigger tiered warnings (such as yellow warnings and red warnings) and push the information to relevant process and quality engineers.
[0094] Step 5: Closed-loop feedback and model self-learning.
[0095] Feedback data collection: The system records detailed information for each warning (factor status, predicted risk value, warning time, and personnel handling the warning) as well as the final quality inspection results for that flight.
[0096] Incremental model updates: Regularly (e.g., weekly), new "alert-result" feedback pairs are used as incremental data to retrain model parameters, especially for tuning. and This makes the model's predictions increasingly closer to actual production conditions, forming a continuously optimized closed-loop learning system.
[0097] To quantitatively evaluate the effectiveness of this prediction method, the following set of comprehensive indicators are used for performance assessment:
[0098] Precision accuracy:
[0099]
[0100] in, (TruePositive) represents the number of correct alerts. (FalsePositive) represents the number of false alarms. This metric measures the accuracy of the warning.
[0101] Warning recall rate:
[0102]
[0103] in, (FalseNegative) represents the number of missed reports. This metric measures the system's coverage of real-world risk events.
[0104] Risk Score RMSE:
[0105]
[0106] in, It is a binary true label (1 indicates that a defect has occurred, and 0 indicates that it has not occurred). This is the probability of a defect occurring as predicted by the model. This metric measures the accuracy of the model's prediction of risk probability.
[0107] Compared with the traditional quality inspection model that does not incorporate this method, the application of this method improves the accuracy of early warning by more than 25% and reduces the false alarm rate (i.e., 1-Recall) by more than 40%, effectively enabling early detection and intervention of quality risks and avoiding the occurrence of major assembly quality problems.
[0108] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any modifications or substitutions made by those skilled in the art within the scope of the technology disclosed in this application should be covered within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.
Claims
1. A method for predicting the frequency of quality problems in aviation assembly processes based on multi-factor risk prediction, characterized in that, Includes the following steps: S1 constructs a many-to-many mapping matrix to digitize and computably represent the relationship between factors and defects, including: S11 defines a structured factor library, which, based on aviation assembly processes, precisely defines all potential risk factors from six dimensions: personnel (U), equipment (E), materials (M), methods (W), environment (S), and monitoring (I), forming a structured data set that can be identified and processed by computer systems. S12 constructs a mathematical matrix Used to store the correlation between risk factors and quality defects, matrix elements Representing the The factor affects the first The initial impact weight of each defect; S13 quantifies factor attributes, assigning three quantifiable technical attributes to each factor: frequency of occurrence. Severity Detectability This provides a data foundation for subsequent quantitative calculation of risk values; S2 constructs a dynamic risk prediction model that integrates data and knowledge. It effectively merges industrial system data with expert knowledge and outputs quantified risk values through interpretable algorithms, including: Data preparation involves collecting historical assembly data, quality records, and process parameters; constructing training and testing sets; and preprocessing the collected data. Through standardized data interfaces, historical process data and final quality inspection results corresponding to the structured factor library are automatically extracted from the manufacturing execution system, quality management system, and shop floor sensor network. ; The extracted data undergoes cleaning, alignment, and normalization preprocessing. It is then sliced and labeled according to assembly frame, workstation, and work step, ultimately constructing a structured training dataset suitable for machine learning models. ,in It is a risk factor state vector, which represents the state of each risk factor in this assembly process; Structured training dataset The training sets are divided chronologically into sets used for model training, parameter tuning, and model selection. and the test set used for final performance evaluation ; The S3 training base probability prediction model uses a gradient boosting decision tree as the base learner. Learn a vector from the risk factor state for input. Probability of occurrence of underlying quality defects mapping function This means learning purely statistical correlations from data; and continuously optimizing parameters to reduce the probability of defects occurring. Approaching the final quality inspection result in an infinitely close manner In a statistical sense ; S4 introduces a dynamic weight fusion module, which employs an induced ordered weighted average operator. Using the historical prediction accuracy of expert experience weights and data-driven weights as the "inducible value," it dynamically generates the optimal weighted vector. This dynamic weight fusion module receives expert experience weights from the many-to-many mapping matrix in S1. And data-driven weights derived from the feature importance of gradient boosting decision tree-based probabilistic prediction models in S3. By inducing an ordered weighted average operator for fusion, the final optimized weights are output. ; S5 calculates the final risk using a risk value quantification engine that integrates FMEA logic. The core calculation formula is as follows: Indicates defects The overall risk value, that is, the defect under the current factor state. The overall risk level, with a range of positive real numbers, is used for early warning judgment; Representation factor The frequency of occurrence, i.e., the factor The frequency level in the historical assembly process ranges from 1 to 10, with a higher value indicating more frequent occurrence. The data is based on historical data statistics. Indicates defects The severity, i.e., the defect The severity level of the impact on product functionality and safety ranges from 1 to 10, with higher values indicating greater severity. Representation factor The detectability, that is, the ability of existing detection methods to promptly detect and intercept factors. The difficulty level is categorized into 1 to 10, with higher values indicating greater difficulty in detection. The probability of defect occurrence predicted by the base probability prediction model is expressed in the factor state vector. The probability of a defect occurring is given below, with a value range of [0,1]. This indicates the final optimization weight factor for defects. The fusion weights are output by the induced ordered weighted average operator and take values in the range of [0,1]. For the factor in The state vector at time t, describing The state of all risk factors at any given time, with values ranging from a multi-dimensional vector, is derived from real-time data from the manufacturing execution system, quality management system, and workshop sensor network. when When the preset threshold is exceeded, an early warning message will be automatically generated and pushed. S6 deployment; S7 self-learning and closed-loop update.
2. The method for predicting the frequency of quality problems in the aviation assembly process based on multi-factor risk prediction according to claim 1, characterized in that, The S6 deployment specifically includes: The trained dynamic risk prediction model is encapsulated as an independent RESTful API microservice, deployed on a production server, and integrated with the manufacturing execution system to continuously monitor or periodically pull real-time assembly data streams; for each set of received real-time factor status data... The service completes risk assessment within seconds. The calculation.
3. The method for predicting the frequency of quality problems in the aviation assembly process based on multi-factor risk prediction according to claim 1, characterized in that, The S7 self-learning and closed-loop update specifically include: The system automatically associates the context of each warning with the final quality inspection result label of the production task, forming a structured "warning-result feedback pair"; it periodically initiates an automated model update task, using the accumulated feedback pairs as incremental data to retrain the base probability prediction model and refresh the final optimized weights. ; The context includes the risk factor status, predicted value, and timestamp.